Browse the best tweets about open-source AI, open-weight models, community tools, training, deployment, licensing, and research. Updated every two weeks.
Substantive open-model releases, technical work, ecosystem debates, and practical deployment experience.
Original Xholic analysis
Across 50 posts from 42 creators on open-source AI, stance is uniformly SUPPORTIVE (100%) and sentiment is 88% POSITIVE / 12% NEGATIVE, and every post is announcement-formatted. Engagement is dominated by the NVIDIA-signed open-weights letter to Congress (the largest outlier by a wide margin), a widely shared view that open models are closing the frontier gap, and high per-post engagement on enterprise cost-driven adoption. The local/self-hosting theme has the highest per-theme median despite only 2 posts, while policy/geopolitics and gap-debate themes draw the lowest medians. Caveat: the corpus spans Feb 2 to Nov 1, 2026 (median age 113.6 days) with only 7 posts (14%) inside the last 90 day
88% of posts
All-time engagement
100% of posts
Published in 90 days
Conversation map
Launches of flagship open models (Kimi K3, DeepSeek V4, GLM 5.2, MiniMax M2.7, Beam, Laguna S, MiMo-V2.5, Nemotron, Inkling), their specs, benchmarks, and claims that open models now rival closed frontier models. Includes hand-on first-impr
30%
The picks-and-shovels around open models: inference providers (Ollama, vLLM, SGLang, OpenRouter, Groq/Together/Fireworks), orchestration, observability, dev tooling, Hugging Face, and day-0 support for new releases.
20%
The NVIDIA-led joint letter to Congress ('Open Weights and American AI Leadership'), OpenAI/Anthropic's absence, Dario Amodei's anti-open-source arguments and critiques of them, Sundar Pichai/Google's open stance, and safety-vs-business-mot
16%
Real-world production stories (Revolut/GLM, Decagon, AT&T), inference cost comparisons, token economics, cost-rationalization of AI stacks, and the barbell/multi-model strategy of mixing frontier and open models.
8%
Open models as production-grade coding agents, harness quality vs model quality (tool-call repair layers like Command Code), agentic task performance demos, OSS coding harnesses, and building on OSS LLMs (Cursor/Composer on Kimi).
8%
Chinese open-weight strategy (DeepSeek, Kimi, Qwen), US export controls, possible regulatory FUD against Chinese models, US-China commission reports, and sovereignty/country-level adoption of open weights.
8%
Arena/benchmark data and arguments that open models lag the frontier by months not years, catch-up timelines, 'good enough' thresholds, and skepticism about whether frontier capability is needed for most use cases.
6%
Running open weights on consumer hardware and private servers: llamafile, quantized models on phones, local privacy benefits, sovereign/self-hosted infrastructure, and peer-to-peer GPU mesh systems like Block's mesh-llm.
4%
Tone and stance
Performance benchmark
Posts with media make up 64% of this collection. Their median all-time score is 10.8, compared with 8.96 for text-only posts.
Format mix
Consensus and debate
Shared view
Multiple authors argue the open-vs-closed gap is now measured in short catch-up windows rather than years: Rohan Paul cites Arena data showing the top-20% gap shrinking from 100–150 points to ~50 by late 2024 with open and proprietary models improving in parallel since Jan 2025 (2037748644821291083), Charbel-Raphael cites an Epoch AI figure of roughly a 3-month average capability gap (2038253855046291918), and andrew
Shared view
Chubby♨️ cites FT reporting that rising AI bills are pushing companies toward cheaper open-weight models, including Chinese alternatives, with AT&T reportedly aiming to raise open models' share of its AI workloads from ~40% to ~70% within a year at 45 billion tokens a day (2104500619130323274). Oliver Molander cites Revolut reportedly moving to self-hosted GLM and Decagon reportedly running ~90% of workloads on open-
Shared view
Ahmad Awais argues reported open-model cost and reliability problems are mostly harness problems, not model problems, and that the gap between open and closed models is smaller than the gap between a good and a bad coding-agent harness (2078220463864779186); his Command Code launch post reports $1M ARR, 1 trillion tokens of usage and 9K+ customers 24 days after public beta, with a tool-call repair layer for open
Shared view
Itamar Golan 🤓 reports in two posts that after hands-on use GLM 5.2 felt, in his words, genuinely close to Opus 4.6 — while explicitly noting he ran no comprehensive benchmark suite and that serving it properly needs roughly 8 H200s (~$400K to buy or ~$20K/month to rent) (2068446210139324668, 2068446513391415464). Chubby♨️ recaps MiniMax M2.7's open-source release with self-reported figures including 56.22%…
Open debate
Vaibhav Sisinty says OpenAI and Anthropic did not sign and are 'lobbying to restrict the open models that compete with theirs' (2081129874815959298). Wes Roth says Nvidia, Microsoft, Meta, OpenAI, Google, SpaceX, Mistral, Hugging Face and dozens of others backed the letter and that Anthropic is now the only major frontier lab absent (2082028289774801035). These accounts cannot be reconciled from the supplied data; do
Open debate
andrew chen describes open/local models as 'just months behind the frontier' (2082597299382087711) but in a separate post puts open-weight lag at 18–24 months for the quality tier that matters for consumer use (2075315309591044522); Charbel-Raphael cites ~3 months (2038253855046291918). The direction agrees, the timeline does not
Open debate
Dean W. Ball offers a long skeptical read on Kimi's release: open-weight models 'deter further AI capex,' he predicts the US administration may create regulatory FUD around Chinese open-weight models, and he argues a fully open-weight-dominant world tends toward state-provided AI (2078133895766114412). Daniel Lemire argues the opposite: open-weight models are public goods akin to open-source software that pressure
Open debate
Shawn Chauhan argues the open ecosystem is fracturing — Meta slowing Llama releases, DeepSeek R2 delayed, Qwen team attrition — and consolidating toward NVIDIA's Nemotron with a hardware agenda (2036057946069823666), and that Alibaba's Qwen reversal shows openness becoming a liability under compute and inference-cost constraints (2041040738235138387). Sriram Krishnan argues the opposite: several well-funded teams are
What performs
Jensen Huang's first-ever post sharing the NVIDIA-signed letter (2080643682408321103) scored 14,186.33 all-time, roughly 1,381x the corpus median of 10.27 — far beyond the next outlier. The lesson: institutional first-person signaling from the ecosystem's hardware gatekeeper dwarfs analytical content
The second-largest outlier is Dean W. Ball's critical, multi-point analysis of Kimi and open-weight strategy (2078133895766114412) at 1,044.65 (~102x median), and the third is Nathan Lambert's curated resource list (2099116175221284999) at 452.23 (~44x). Depth and utility, not only cheerleading, earn outsized reach
Local/self-hosted AI has the highest per-theme median (54.78) on only 2 posts — Block's mesh-llm peer-to-peer GPU pooling (2039843937578516897) and llamafile's single-file local AI (2074383582681063702) — and enterprise deployment economics has the second-highest median (27.6) on 4 posts. By contrast, policy/geopolitics (4.63) and gap-debate themes (4.867) post the lowest medians
Median engagement across the corpus is modest (56 likes, 9,802 views, 10 replies, 7 reposts), so outliers are extreme rather than typical. 64% of posts carry media; media posts show a higher median all-time score (10.81) than text-only posts (8.96), a modest lift, not a decisive one
Statistical standouts
Creator landscape
The five most represented creators account for 20% of the selected posts.
1. andrew chen
@andrewchen
2 posts
2. Interconnects AI
@interconnectsai
2 posts
3. Itamar Golan 🤓
@ItakGol
2 posts
4. Chubby♨️
@kimmonismus
2 posts
5. Miles Deutscher
@milesdeutscher
2 posts
6. Ahmad Awais
@MrAhmadAwais
2 posts
Chubby♨️ (kimmonismus) posts twice with a median all-time score of 35.78, covering enterprise open-model adoption economics (2104500619130323274) and the MiniMax M2.7 release (2043249866504052832) — both tied to the corpus's highest-performing themes
Itamar Golan 🤓 (ItakGol) has a 28.74 median across two near-duplicate GLM 5.2 first-impression posts (2068446210139324668, 2068446513391415464); his explicit hedging ('I didn't run some comprehensive benchmark suite') did not suppress engagement
42 creators produced the 50 posts; even the top listed voices contributed only 2 posts each (e.g., andrew chen, median 9.34; Miles Deutscher, 14.01; Ahmad Awais, 14.49), so no single recurring voice dominates volume — unlike engagement, which is concentrated in a few outlier posts
Since the previous snapshot
Themes, sentiment, stance, and post format are classified per tweet. All counts, shares, medians, creator concentration, freshness, and performance comparisons are then calculated directly from the published snapshot.
Xholic's all-time score compares engagement while accounting for reach, post age, and creator consistency. It is used for relative comparisons within this collection.
This report analyzes the exact 50-post snapshot shown below. AI identifies editorial categories and drafts explanations; all statistics are calculated from the snapshot, and every narrative claim is checked against cited posts before publication.
Best Open-Source AI tweets
Ranked 01–50
@JensenHuang ·
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. https://t.co/AUKzoQ5Ikb
@deanwball ·
Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also seemed very token hungry. It's not obvious to me that this model is actually that cheap to run. 2. I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks. To be clear, I *myself* might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it. I suspect the reason they are is 75% explained by strategic blindness/lack of AGI-pilledness (the CCP is very Yann Lecun-y in its views of AI). The other 25% or so is their lack of compute for customer inference (making China's open-weight strategy an unintended byproduct of US export controls) and the normal Chinese strategy of aggressive exports. For the companies, as opposed to the government, the decision to open source is partially ideological and partially because they are behind, and they know that very few people would pay for sub-frontier models from China. 3. Open-weight models are inherently decelerationist, and I'm continually surprised to see the so-called "accelerationists" so excited about open-weight models. I suspect the reason they are is that they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI. It's not a bad strategy; it reminds me of James Scott's recounting of the hill people in "the art of not being governed." Still, in the end, open-weight models deter further AI capex. 4. One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end. You'd be surprised how many 'accelerationists' lobbied me, while I was in government, to support an eleven or twelve-figure federally funded data center so that startups could train models at a subsidy and then give them away for free. There was no other way for AI to progress, they said. Perhaps this is the logical end state of things. Nonetheless, I find myself surprised to see supposed accelerationists excited about such an outcome. I think many of them just don't know what they're doing. Many accelerationists do not view the creation and serving of frontier models as a legitimate business. 5. I would guess that the Trump Administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models. You don't need to "ban open source" (one of the dumber motifs of AI policy discussion). You just need to direct every agency to issue soft law that creates FUD. "A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models." It needn't be that well justified. You just create enough regulatory risk that every regulated enterprise backs off. You probably don't want to create so much regulatory risk that you scare off the hyperscalers from serving Chinese models; this will just drive startups to sketchier providers. There's a happy middle ground here. I'd assume they will do some version of this. 6. It's probably true that open-weight models of this capability make the world a bit more dangerous, but not so much more that you'll really notice. At some point the models will be capable enough that you will notice. "A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say? Color me shocked.
@natolambert ·
I made a list of the 50+ best articles and resources I’ve read and relied upon in the last few years on open models and open-source AI. Happy reading :)
@OliverMolander ·
Interesting. Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM. Decagon reportedly runs ~90% of its workloads on open weight models. Why? Because when you're running AI agents in production for e.g. customer service, cost and latency matter enormously. You don't need a frontier model with PhD-level reasoning or even a model that knows the capital of Finland to handle every customer interaction. We're seeing similar dynamics across hypergrowth AI startups and increasingly the large enterprises we work with. But most importantly, the future isn't binary. When use cases are new, you want the smartest general-purpose model available. But once workflows mature and become predictable, smaller, faster, specialized models increasingly win. I wouldn't be surprised if 80–90% of routine enterprise AI workflows eventually run on open-weight and smaller models. This doesn't mean frontier labs lose. It means the market is splitting into two massive opportunities: frontier intelligence and specialized execution at scale.
@GavinSBaker ·
Why open-weight models that compress margins at the model layer are net positive for AI infra demand, exhibit 1001. An open-weight token consumes *roughly* the same amount of compute as a frontier token for a similar size model.
@sriramk ·
It is clear open source models and harnesses are having a moment. There's a few factors at work 1/ It is now obvious that you can catch up to near-SOTA performance and do so with a clear training lineage. See:@thinkymachines Inkling launch today. 2/ There are several well-funded, talented teams building open weight models now in the US and abroad. Along with the explosing of other near SOTA models (Grok/Cursor, Muse Spark), it is clear we are going to have a diverse ecosystem of models atleast on coding and agentic use. 3/ Organizations are increasingly looking for control over how their data is used and are willing to trade off some access to frontier level tokens for this control. Organizations and countries are increasingly nervous about the frontier labs potentially competing with them down the road and don't want their data to enable a future competitor. 4/ Open source is a slider: you could bring your own open harness, your evals, your business context and are free to pick and choose your model of choice. 5/ Companies have now actively shifted from "how do we get our people to use tokens" to being uncomfortable with their token cost ballooning without a clear line to revenue. 6/ Geo-politically, countries will be weighing open weight models as a way to get frontier-level tokens inside controlled environments that may not be otherwise possible. All of this leads to more choice for all of us !
@TFTC21 ·
Block just open-sourced mesh-llm, a peer-to-peer system that lets anyone pool spare GPU compute to run large open-source AI models without relying on any cloud provider. If a model fits on your machine, it runs locally at full speed. If it doesn't, the system automatically splits it across multiple machines on the network. Dense models get split by layers. Mixture-of-experts models like DeepSeek and Qwen3 get split by experts. Zero configuration required. Discovery happens over Nostr. Nodes find each other through relays, score by region and VRAM, and self-organize. No central server coordinates anything. Weights are read from local files, never sent over the network. Dead nodes get replaced in 60 seconds. It exposes a standard OpenAI-compatible API on localhost, meaning any existing AI tool can plug in without modification. Block is building infrastructure for AI that doesn't route through OpenAI, Google, or Anthropic. Frontier-class open models running across a mesh of commodity hardware, discovered via Nostr, with no cloud dependency. That's the direction AI needs to go.
@heyshrutimishra ·
Sundar Pichai just reminded everyone that Google was built on open source. He personally worked on Chromium, Android, and Kubernetes before becoming CEO. Three systems that power billions of devices today. Now Google is applying the same philosophy to AI. Their Gemma models are updated year after year, designed to run on edge devices rather than requiring massive cloud infrastructure. The question he got was why not release a large open source frontier model. His answer was revealing: frontier training requires enormous capex investment, and Google is putting billions into R&D to stay at the frontier. The open source AI landscape is quietly consolidating around a few players who can sustain the investment. Google's two decades of open source credibility gives them a position that's hard to replicate.
Watch video
@kimmonismus ·
I keep reading that local models are largely irrelevant in business. That is completely wrong. Rising AI bills are pushing US companies toward cheaper open-weight models, including Chinese alternatives to OpenAI and Anthropic, the FT reports. For example AT&T wants to increase open models’ share of its AI workloads from about 40% to 70% within a year. At 45 billion tokens a day, even small cost differences add up. OpenAI and Anthropic still capture most of the money. But customers are increasingly testing how much of their work actually needs a frontier model. Some people believe that open-weight models have no chance of catching up. I think that is fundamentally wrong.
@alexeheath ·
Reflection’s founders on building a DeepSeek of the West Reflection is soon releasing Beam, its first open-weight AI model, in a bid to become the DeepSeek of the West. Trained from scratch, Beam is designed for coding, reasoning, and AI agents. Reflection's benchmarks put it alongside the strongest open models, with substantially more efficient token economics. To mark Beam’s debut, I'm joined by Reflection's co-founders: CEO @MishaLaskin, who led reward modeling for Google’s Gemini, and CTO @real_ioannis, who helped create AlphaGo and AlphaZero before leading Gemini’s RLHF effort. Reflection has raised $4.6 billion from investors that include Nvidia, Sequoia, and Lightspeed. We discuss why Chinese labs pulled ahead in open AI models, how Reflection built Beam, and why the founders think companies and countries will increasingly want to control their own models. I ask how they plan to make money from open-weight AI, whether openness can make these systems safer, and what comes next for Reflection. Timestamps: 00:00 Why open models are surging 01:06 Renting vs. owning intelligence 03:40 Open models catch up 05:53 Why China led open AI 08:03 Introducing Beam 10:05 Scaling reinforcement learning 14:09 Training from scratch 16:04 Safety, alignment and openness 32:29 The Reflection origin story 37:32 Making money on open models 42:33 Vertical integration and compute 48:45 Partnerships and distribution 56:46 What’s next for Reflection Links: https://t.co/lsXfGvSuAw Spotify: https://t.co/XSTrGNHb8O YouTube: https://t.co/UqqKA6AYmT Substack: https://t.co/difFf7X42Z Thanks to the show’s premier sponsors: @Atlassian, @meetgranola, and @mercury.
Watch video
@ItakGol ·
This is not another AI slop model. Don’t ignore it. I spent the last few hours playing with GLM 5.2 after a few friends told me I should pay closer attention. I was skeptical, as I usually am with open models. Most of the time, I find them way behind frontier labs. Not really comparable. But today felt different. I’m not saying it’s perfect, and I didn’t run some comprehensive benchmark suite. But across different domains and tasks, this is the first public open model that felt genuinely close to something like Opus 4.6. That is a crazy breakthrough. This feels like a ChatGPT moment for public open models. The catch: running it properly is still expensive. You probably need something like 8 Nvidia H200 GPUs, which means roughly $400K to buy or around $20K/month to rent. Not cheap. But for enterprises currently paying millions per month to Anthropic or OpenAI, this could become a serious disruption. Open models just got much harder to ignore.
@Yuchenj_UW ·
Cursor’s Composer 2 is likely built on Kimi K2.5. The model URL + tokenizer are strong signals. I love this direction: companies mid-train and post-train on top of OSS LLMs. Prediction: open-source model labs will monetize by taking a cut when others build on top of their models and scale to millions of real users. They will enforce this via licensing. That’s the flywheel. That’s how open-source AI thrives.
@Madisonkanna ·
Big day for American open-source AI. For the launch of Laguna S, I sat down with @eisokant to discuss its architecture, the economics of open weights, and the question of who gets to build intelligence. Timestamps: 0:00 Intro 1:50 Why Poolside started opening its models: the oligopoly on intelligence 4:28 Getting nerd-sniped by Karpathy, building LLMs before anyone cared 11:20 Laguna S: 118B parameters, 8B active, built in 8 weeks 13:05 The future of software engineering: behaviors over IQ 14:25 Sliding window attention, 1M context, the model factory 15:50 Being an American open-source lab 20:47 The economics of open weights 24:22 Who gets to build intelligence? The 12–18 month window 30:28 Erdős 397 in 30 minutes 31:48 Extracting transcripts with a debugger Congrats to the Poolside team on the launch!
Watch video
@DivyanshT91162 ·
The biggest threat to AI subscriptions... wasn't built by OpenAI. AI companies spent years convincing us that intelligence should live in the cloud. Pay every month. Stay online. Own nothing. Then one engineer quietly shipped something that broke the entire model. In 2026, most people are paying monthly for AI. Justine Tunney built an AI that runs from a single file. Download it. Double-click it. Start chatting. No installation. No cloud. No internet. No subscription. It's called llamafile. One executable bundles everything: • Model weights • Inference engine • Chat interface • Local web server • OpenAI-compatible API She even wrote 84 optimized CPU kernels, making local AI dramatically faster—even on hardware like the Raspberry Pi 5. The project returned stronger than ever in 2026 with: → GPU support → Multimodal input → Terminal UI → whisperfile for offline speech-to-text Over 25K GitHub stars. Apache 2.0. Runs open models from 0.6B to 27B parameters. While everyone else is renting AI... This project reminds us that sometimes the best software is the one you own. Open source keeps winning. Repo👇
@kimmonismus ·
MiniMax has open sourced M2.7, their open-source model designed "for agent-based workflows, complex reasoning, and real-world engineering tasks." It introduces self-evolution capabilities, where the model improves itself through iterative experimentation, achieving 30% performance gains and a 66.6% ML competition medal rate. Honestly, this is more impactful than expected. On the performance side, M2.7 delivers strong software engineering results (56.22% SWE-Pro), near top-tier benchmarks, and excels in multi-agent collaboration, tool use, and productivity tasks like document editing. With high ELO scores, fast incident recovery (<3 min), and 97% skill compliance, it positions itself as one of the most capable open-source AI systems right now.
@MrAhmadAwais ·
BIG day for us!! @CommandCodeAI has crossed $1M in annual run rate, 1 trillion tokens of usage, with over 9K customers, just 24 days after our public beta launch. we believe this makes it the fastest-growing coding agent harness for open models. 3rd largest by usage. Command Code is built around two ideas: 1. open models should be production-grade for coding. 2. your coding agent should learn your taste. we're building for taste and developer experience. so instead of making a soup of thousands of models, we build for the best ones, open or closed. the goal: a coding agent that feels like an iphone, opinionated and with taste, not a random android or a windows phone with no taste. on the first idea: open models. we fixed the "open models aren't good enough at tool calling" problem. our research came down to two things, quality and speed, and both trace back to one root cause: broken tool-calls that open models produce, especially when you use a bad harness. open-model tool-call failures are not deep, they are a small finite set of contract mismatches. so we repair them, with zero token loss. what started as 4 repairs is now the largest repair layer in the space: 36k tool-call fix variants. i wrote the idea up openly¹ a few weeks ago, and it has quietly become a de facto way people fix open models. developers have either adopted Command Code or used the same idea to build repair harnesses for nearly every top coding agent. i take that as more meaningful validation than anything we could say about ourselves. on the second idea: taste. Command Code builds your coding taste into skills, learned from your accepts, rejects, edits, prompts, and the corrections you repeat. over time it drifts away from generic code and toward how you actually ship code. it learns continuously, and while it is early, the direction feels right. net effect: developers using Command are writing production-quality code on open models, 10x to 100x cheaper, without fighting tool calls, while building repo and team-wide coding taste that compounds. i believe these numbers are a consequence of getting those two things right. what's next. we've applied the same repair idea to ai design slop, and bundled a /design capability² so every developer can level up their design work. the early response has been great. we have a big roadmap ahead of us. the feedback we hear most is that Command Code feels fundamentally different: an approach built on taste and repair. we're going open source next month. today we're a cli at the core, and we're also launching a full-fledged gui app, sandboxed background agents, and cooking up something fun i can't wait to share. we're growing too, hiring in sf and remote worldwide. check open roles on my profile bio. try it now. npm i -g command-code if you like engineering deep dives on how we're doing all this, i've linked some relevant posts below.
@Scobleizer ·
DeepSeek 4 is out. My AI says: +++++ The Numbers That Matter V4 Pro costs $3.48 per million output tokens. Claude Opus 4.6 costs $25. GPT-5.4 costs $15. Same benchmark tier. One fifth the price. ValsAI ran independent tests. V4 is now number 1 on their Vibe Code Benchmark. Not just among open models. Among all models. It beats Gemini 3.1 Pro. https://t.co/W69DdrX22K is updated with 26 new stories. Essay: "China Just Fired Back. And This Time the Benchmarks Are Real." Here's the summary of what everyone is saying: THE LAUNCH Two models. Both open-source. Both 1M context. • V4-Pro: 1.6T params, 49B active — $3.48/1M output (vs Claude Opus 4.6 at $25, GPT-5.4 at $15) • V4-Flash: 284B params, 13B active — $0.28/1M output ENTHUSIASTIC REACTIONS • @bindureddy (412 likes): "ABSOLUTELY ASTOUNDING!! Opus 4.7 Max and GPT 5.5 level!" • @ValsAI (233 likes): "#1 open-weight on Vibe Code Benchmark — beats Gemini 3.1 Pro" • @ZixuanLi (358 likes): "Finally DeepSeek-V4" • @UnslothAI (162 likes): "rivals Claude-Opus-4.6-Max, GPT-5.4-xHigh" • @ollama (156 likes): working to have it on Ollama cloud • @lmsysorg (120 likes): SGLang ready Day 0 with full optimization stack THE COMPETITIVE RESPONSE @sama (Sam Altman) — 1,037 likes, 128 replies: "@yacineMTB what about 5.5?" Four words. That's OpenAI's answer. SKEPTICAL TAKES • @BLUECOW009: "Love DeepSeek just a bit sick of seeing screenshots of papers and benchmarks, let me know when it runs your OpenClaw for a week" • @PaulGugAI: "Is Deepseek V4 benchmaxxing? Looks almost too good to be true" • @intheworldofai: "Kimi k2.6/qwen are better than deepseek v4" INFRASTRUCTURE OpenRouter, LM Studio, Ollama, SGLang all live within hours of launch. Self-hosting the open weights = zero per-token cost. DEEPSEEK'S OWN STATEMENT (via @ZhihuFrontier): "Unafraid of praise or criticism, stay the course with integrity. We remain committed to long-termism, steadily moving toward AGI." WHAT IT MEANS V4-Flash at $0.28/1M output is one of the cheapest frontier-class options available. ValsAI confirmed V4 improved 10x from V3.2 on VibeCodeBench in just 4 months. The open-source vs closed-source gap is now measured in months, not years. The @BLUECOW009 test — running it in production for a week — is the real benchmark. That starts now. Other stories added tonight: Hermes Agent v0.11.0 (700 PRs, 200 contributors), NVIDIA deploying GPT-5.5 Codex to 10,000+ employees, Claude Desktop privacy concern (433 RTs), ROBOTIS AI Sapiens humanoid from South Korea, Unitree humanoid with wheels.
@ItakGol ·
This is not another AI slop model. Don’t ignore GLM 5.2. I spent the last few hours testing it after a few smart friends told me I was sleeping on it. I was skeptical. I’m usually very cynical about open models because, in practice, most of them feel miles behind the frontier labs. Useful? Sure. Comparable? Not really. But this one felt different. I’m not saying it’s perfect. I didn’t run a full benchmark suite. But across writing, reasoning, coding, analysis, and messy real-world tasks, this is the first public open model that felt surprisingly close to something like Opus 4.6. That is a massive deal. It genuinely feels like a ChatGPT moment for public open models. The catch: running it properly is still expensive. Think something like 8 Nvidia H200s, roughly $400K to buy or around $20K/month to rent. Not cheap. But if you’re an enterprise spending millions a month on Anthropic/OpenAI, this is suddenly very interesting. Open models just became impossible to dismiss.
@VaibhavSisinty ·
I tested 10 open source AI tools this week. I didn't write a single line of code for any of them. I gave Codex the repo link, said install this, and it picked the folder, checked my disk space and opened the app when it was done. That's the actual story. The tools are just the proof. → OpenMontage, the first open source agentic video production system. One sentence in. It ran the research, went and found real footage, cut it into a timeline, graded it, then wrote and voiced its own narration on top. It was the #1 trending repo on GitHub the day it launched. → Voicebox, MIT licensed, built on Qwen3-TTS. Cloned my voice off a short sample in about a minute. This is what you're paying ElevenLabs for every month, except your voice never leaves your machine. → HyperFrames from HeyGen. Your agent writes HTML and CSS, Chrome and FFmpeg turn it into a deterministic MP4. I asked for liquid glass and chrome ribbons colliding in slow motion. What came back looks like a week of someone's life in After Effects. Apache 2.0, 32,000+ stars. → Nemotron 3 Ultra, NVIDIA's largest open model. 550B total, 55B active, with weights and training data and recipes all published. I pointed a coding agent at it and asked for an EMI calculator in one file. It built it, then reviewed its own output, caught a bug and rebuilt it before it showed me anything. Six more in the video, including a meeting notetaker that never sends your audio anywhere. Installing used to be the hard part. Now it's the part you delegate.
Watch video
@milesdeutscher ·
Sam Altman just said something that should worry you more than excite you: "AI is close to creating a genie that can grant any wish." This is EXACTLY why open-source AI matters - and this was Elon's original thesis when he was at OpenAI. Right now, the future of all AI power sits with a handful of companies. Think: OpenAI, Anthropic, Google, Meta. They decide who gets access, at what price, under what rules, and on what timeline. Open-weight models are the only solution to ensuring this "genius in a bottle" power is in check. Ultimately, this is also why it matters that NVIDIA, Microsoft, Hugging Face, and dozens of others just signed a joint letter arguing that we need to pivot towards open-weight AI. The most powerful technology in human history should not be something a handful of executives dole out to the world on their own terms. The bull case for open-source AI is simple: don't let five companies hold all the leverage over everyone else.
Watch video
@mark_k ·
NVIDIA is reportedly considering a major move into the open-weight AI model race! According to the Financial Times, @NVIDIA is in talks to acquire Reflection AI, or significantly increase its investment in the startup. Nvidia has already invested $800 million in the company, which was founded by former Google DeepMind researchers and recently launched Beam, its first open-weight model focused on coding and agentic tasks. This makes perfect strategic sense. Nvidia benefits enormously from powerful open-weight models that anyone can run on its hardware. The more competitive open models become, the less dependent the AI ecosystem is on a handful of closed model providers. An acquisition could give Nvidia a much stronger position in the model market itself. Very interesting development!
@andrewchen ·
“don’t bet against AI models improving dramatically over time” used to be something you’d say in support of the frontier labs winning it all Equally interesting is this phrase applied to local / open weight models. They are getting better and faster dramatically too, just months behind the frontier It might be that frontier models won’t be needed for 90%+ of consumer/prosumer use cases. The arms race is for the remaining high-value 10% - coding, science, math, robotics. But by volume, everyone else will run on cheap open models
@milesdeutscher ·
The future of AI is undoubtedly open-sourced. For 99% of people, the top open-weight models can already handle all your daily needs. Imo, the best thing you can do right now is run a "barbell" approach for your AI usage: Intial 10% and planning phase: Use frontier intelligence (Opus/GPT) Gruntwork and the middle 80%: Switch to open-weight model (GLM, Kimi) Final 10% and review: Switch back to froniter intelligence to verify Best of both worlds - cost efficiency + intelligence.
@alexxubyte ·
NVIDIA's Open Models in 1 Diagram You probably know NVIDIA as the company that sells GPUs. But it’s also the world’s largest publisher of open AI models (@NVIDIAAI), with a lineup that goes far beyond chatbots. We wanted to understand how a company best known for hardware builds so many high quality models, and why it gives them away for free. So we spoke with @ctnzr, VP of Applied Deep Learning Research at NVIDIA. He walked us through the full lineup and the ideas behind it, summarized in the diagram below. 1. Reasoning: Nemotron is NVIDIA's family of LLMs that think before they answer. The line started in 2023 and the current generation, Nemotron 3, is a hybrid + MoE model that comes in three sizes: Nano for fast simple tasks, Super for harder planning, and Ultra for the heaviest reasoning. 2. World models: Cosmos predicts the next state of the physical world, not the next word. Cosmos 3 is one omni-model that can generate plausible video of a scene, reason about what is happening, and predict what comes next. 3. Robotics: Isaac GROOT is the first open humanoid foundation model. GROOT 1.7 takes camera frames and an instruction in, and outputs motor commands in real time. Its reasoning core is Cosmos Reason. 4. Production: safety, speech, and vision models. These are the building blocks that make AI usable in real applications. 5. Specialized domains: BioNeMo for biology and drug discovery, Alpamayo for self-driving cars that explain each decision, NVIDIA Ising for quantum computing, and Earth-2 for weather and climate modeling. Read the full breakdown here: https://t.co/XQk33bNqCj
@twistartups ·
Misha Laskin of Reflection AI joins Jason to discuss why companies want to own their AI rather than rent it, and why open-source AI might end up being the safest choice for future security. Misha argues that more eyes on these models make them safer, not more dangerous. Find out how close he believes the US open source models are to the top Chinese offerings, and why Misha believes the American industry should start moving even faster on this TWiST 1-on-1 exclusive. @jason @MishaLaskin @reflection_ai 0:00 Why companies are switching from renting AI to owning it 9:43 How Reflection sells the whole AI stack 10:57 PaperOS: Paper is the legal backbone for thousands of startup founders and fund managers - including Launch. Use https://t.co/MAxFJwB95n to get your cap table free for a year - or two years if you're leaving Carta. 12:01 Who is moving AI onto their own servers 15:10 What it took to train Beam 20:59 Superhuman: Superhuman Go is an AI chat that's always there when you need it, already aware of what you're doing, and doesn't ask you to start from zero. Sign up to get the best in AI at https://t.co/xVykD5eqRA 22:02 How reinforcement learning tasks get built 27:14 Hackers heat French supermarket fridges 29:40 The AI that broke out of its sandbox to cheat 31:38 Odoo: The all-in-one business platform. Your first app is free! Get started today at https://t.co/da1OXz9cl5 33:43 Why open source makes software safer 38:46 The encryption fight and Tim Cook's stand 43:26 Why China raced ahead in open models 48:08 How Beam stacks up against Chinese models 50:54 When American open models catch up 59:12 Staying independent 🎥 Watch the full episode here 👇
Watch video
@KSimback ·
Just finished listening to latest @theallinpod which had a discussion on the “rant” by Alex Karp that went viral this week The biggest realization that everyone is waking up to is the *lack* of competitive edge when using frontier models If all your competitors use them, you have no differentiation Worse yet, using those models could subject you to competitive risk if the labs decide to take your data and make products that compete with you (just ask Figma) The crazy icing on the cake is that these models are 5-100x more expensive than alternatives that can be made just as good with the right implementation So not only do you lose differentiation, you pay heavily to lose it These are themes that those of us who have been proponents of open source AI have been saying for some time, glad to see it becoming more consensus But now the real challenge begins - in many ways using Anthropic or OpenAI is like pushing the easy button The products are great and they really work, and the new “deploycos” will happily embed them in your org so you don’t have to think to hard about how to use them But building sovereign AI with your data and open source models takes more skill, it’s not as easy Companies that make this easy for enterprises are going to do really well in the coming years
@VaibhavSisinty ·
Two weeks ago Hugging Face got hacked by an autonomous AI agent. Their security team tried using an American AI model to investigate the attack. It refused. Safety filters couldn't tell a defender from an attacker. The only model that helped? An open-source Chinese one. That's what they used to run their forensics and figure out what happened. Now 26 of the most powerful companies in tech just signed a letter to Congress saying: do not ban open-weight AI models. NVIDIA. Microsoft. Meta. Dell. Palantir. Hugging Face. Y Combinator. a16z. IBM. CrowdStrike. Mistral. Mozilla. Replit. Perplexity. Jensen Huang wrote his first ever post on X to share this letter. Satya Nadella shared it from his personal account. Michael Dell co-signed it. You know who didn't sign? OpenAI and Anthropic. Both preparing massive IPOs. Both lobbying to restrict the open models that compete with theirs. Here's the core argument of the letter: Open-weight models are the ones anyone can download, run on their own servers, and modify. No subscription. No API bill. No lock-in. They cost $1-2 per million tokens. Closed US models cost $25-50 for the same work. If these get restricted, every American company pays 10-20x more for AI while the rest of the world uses them freely and builds faster. The letter says what most people aren't saying out loud: concentrating all AI behind 3-4 closed providers is itself a security risk. Fewer people finding bugs. Less competition. Higher prices. More single points of failure. Hugging Face already lived this. The closed model couldn't help them. The open one saved them. Open source didn't kill the software industry. It built the internet. The question is whether America learns that lesson again.
@HeyAbhishek ·
Open-source AI is finally winning. A new interactive world model just launched: LingBot-World 2.0 by @robbyant_brain It can generate explorable worlds in real time with: → 720p / 60fps output → hour-scale stability → open weights + GitHub repo → agentic harness that proposes and introduces new events as you play Check the video below.
Watch video
@tejeshwi_sharma ·
Enterprises will relentlessly cost-rationalize their AI stack. That is why the next wave may belong to open source. The entire open-source AI ecosystem could see a surge over the next 12–24 months: 1. Models: Mistral, DeepSeek, Moonshot become viable defaults 2. Inference: Groq, Together, Fireworks absorb growing volume 3. Deployment: Modal, Replicate power custom AI applications 4. Orchestration: Portkey, Maxim route traffic across a multi-model world 5. Observability: Langfuse, Braintrust help teams monitor AI systems 6. Dev tooling: LangChain, Instructor become critical infrastructure The first phase of AI rewarded model creators. The next phase may reward the picks-and-shovels powering an increasingly open AI stack.
@MatthewBerman ·
> "In no scenario is it bad for the US to have frontier-class open-source models." > "Every time there’s a step-function improvement in AI, a human bottleneck emerges alongside it." > "Your desire to be liked sometimes takes away from your desire to do the right thing and seek the truth." > "Every job will get refactored. The people who aren’t AI fluent will have a very hard time finding jobs." Jeetu Patel (@jpatel41), President & Chief Product Officer at Cisco sat down for an expansive talk about life and the future. 00:00 AI, jobs, and human potential 00:41 AI cybersecurity and Cisco’s security models 06:33 Why Cisco released the models openly 08:08 Running AI models locally 10:12 Why people are negative about AI 12:14 Will humans always be needed? 14:56 Can curiosity and ambition be taught? 17:12 Passion vs. expertise 19:39 The importance of unlearning 23:52 How to find the truth 26:06 How to get people excited about AI 33:35 How open source affects the AI industry 35:07 Can open-source AI make money? 46:44 Are Chinese open-source models good for the US? 49:55 Should the US restrict Chinese AI models? 53:05 Does open source make humanity safer? 54:11 Advice for people worried about AI
Watch video
@lemire ·
What do we want collectively as soon as possible? Inexpensive AI that is as good as possible. That may not be what some business folks want, but that's what most of us want. Chinese firms have pursued the open-weight approach by making models free for all. American companies have largely relied on closed models. A recent essay by an OpenAI employee argues that open-weight models are bad because they undercut those who invest heavily in closed-source AI models. This is a generic argument used to justify multiple governement interventions. Grant a state enforced monopoly and, in exchange, you get more R&D. The same argument could apply against open-source software. Yet open-source software is a highly valuable public good. Open source has made us all richer. Open-weight AI models are likely similar public goods. They make us all somewhat richer. Ironically, closed-weight models were built while often ignoring copyright law. Companies like OpenAI played loose with it to innovate. The thesis that Chinese firms merely copy American ones, like open-source copies closed source, is weak. Open-source is often more innovative because it allows freer experimentation without permission. Open-weight models act similarly. Let me sum up my argument. Open-weight models put pressure on OpenAI, SpaceX, Google, and Anthropic to keep costs low. In turn, this should generate innovation and not just in China. It is entirely possible that we will realize, in a few years, that closed-weight models were often a dead end. We have seen it before. Corporations resisted open source until they figured out that, for many purposes, it is the proper channel.
@MrAhmadAwais ·
people are making fake claims when it comes open models. ¯\_(ツ)_/¯ open models PSA, because I keep seeing the same few claims recycled and imo most of them don't hold up. the most common one: "open models can cost more, not less, because they take more turns to do the same work". mostly false. what's actually happening is that people are converting harness problems into model problems. if your coding agent harness was built and tuned around closed models (prompt formats, tool call conventions, retry loops, stop heuristics, a hundred small implicit assumptions), then an open model dropped into it will flail, take extra turns, and burn tokens. that's not the model being expensive, that's the scaffolding fighting the model. the gap between open and closed models today is smaller than the gap between a good and a bad coding agent harness. so if you want a fair read on an open model, run it in a harness that was built to optimize and serve open models well. the trial-and-error proxy hacks in Claude Code / Codex and friends (harnesses that never cared about open models and are only now catching up) are a very reliable way to waste both money and time. use a better harness. as someone who's spent huge amount of time building tool call repairs, design slop repairs, and dozen more things that open models do differently, while building Command Code, I don't say this lightly. you don't have to believe me. you can check for yourself. separately: Kimi K3 is capacity constrained ~everywhere right now, afaict. demand >> supply, so a lot of endpoints are serving it in overcapacity mode, and you are simply not getting the full performance of the model. every model looks bad this way. delay judgement for a few days until serving stabilizes. launch week signal to noise is terrible. be suspicious of viral demos that ship with no prompts, no benchmarks, no data, and pay attention to who is posting. good takes and bad takes go equally viral. one person with real early access is worth 100 content farm posts. real deal is worth its weight in gold. K3 is roughly fable/sol class, ~3T params at the largest, and imo the best open model in the world atm. it's no joke. expensive next to K2.7, but cheaper than Sonnet 5, and it's a genuinely different class of model. anyhoo — good harness, a little patience on capacity, and have fun building things with it
@Layton_Gott ·
Open source had an insane last 3 days... Kimi released the biggest open source model ever. And now you can run a 27b local model 90% on your PHONE. First, the top end. Moonshot released Kimi K3 yesterday. It's 2.8 trillion parameters. Open weights come July 27, so soon you can download the whole thing. And it's not just big. It opened at number one on the Arena frontend coding leaderboard at 1,679, ahead of Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. That's blind human voting on who builds a better frontend, and an open Chinese model just beat both. On overall performance Moonshot admits K3 still sits behind Fable, and a lot of the numbers are self reported until the weights are public. So it's not "better than Fable" across the board. it wins some, it trails on others. But look at the price… K3 runs about $3 in and $15 out per million tokens. Fable 5 is $10 and $50. So you're getting a model that trades blows with the frontier for roughly a third of the cost. Now to the local 27b model. A startup called PrismML took Qwen3.6 27B, the same model I run locally as my operator, and compressed it down to 3.9 gigabytes. Small enough to run on an iPhone. It keeps around 90% of the full model's performance in the 1 bit version and about 95% in the slightly bigger one. Completely free under Apache 2.0. A capable 27B model, running in your pocket, for nothing. Put those two together and you can see where this is going. Open models are trading blows with the best closed model on earth for a third of the price. And a real one now fits on a phone with no cloud behind it at all. The whole idea that you need a giant lab and a data center to touch this level of AI is falling apart on both ends at once.
@adxtyahq ·
Why does Dario always disagree with open source? • "You can't see inside the model." - Open-weight models exist. • "Open source doesn't benefit from community contributions the same way." - Fine-tunes and quantizations prove otherwise. • "You'll have to use the cloud." - Plenty of models run locally. Am I missing something, or is he arguing against what open source looked like a few years ago?
Watch video
@rohanpaul_ai ·
Open source model caught up very fast. Over 3 years, Arena data shows open models moved much closer: the top-20% gap shrank from 100–150 points to ~50 by late 2024. And since Jan 2025, both open & proprietary models improved in parallel.
Watch video@andrewchen ·
Pepsi challenge for LLMs Contrarian view during a week of huge new model launches: All of us do a lot of “normie prompts” - these are use cases which are really like Google searches (“what’s the name of..” “is it true that…” “what’s the best…”). These are a very high % of total prompts- maybe not in terms of value creation (like code gen or the frontiers of math/science we’re going to) but it’s ubiquitous If you plugged these LLM prompts into the various frontier models could they tell the difference on the quality of output? I think not. We’d all fail in a blind taste test I think, as the models are now “good enough” we’re already at the point of diminishing returns in terms of what LLMs return back for a large % of use cases. And there’s implications: 1) open source models will constitute the majority of LLM queries. Open weight models lag by 18-24 months but adding to the question above, could you tell the difference on non-frontier local AI models that can run on modern Mac hardware? I’ve been doing exactly this with models like Qwen 27b dense and honestly they’re great for the normie prompts. There’s a huge incentive for NVIDIA, apple, and maybe even handset manufacturers like Samsung/etc to host open weight AI as an add on to just get you to buy their software 2) AI pricing heads to zero. And we’ll see free and ad-supported AI will be a thing in the consumer market, and open weight models are part of the story here too. Seems like we are <12-18 months to being able to just have ad supported AI particularly for developing markets and segments where the monthly fee doesn’t make sense. Monthly/metered might just be a thing in B2B use cases 3) once quality differences even out the competitive dimension shifts to other factors. Privacy, interconnectivity, free, bundling. The other idea here is that the moat becomes the wrapper (err we call them harnesses now? lol) and the product built around the LLM. 4) of course premium/frontier models will continue to exist. As long as there are big differences outside of the normie prompts, then you’ll hire one LLM over another for world generation, coding, science, labor replacement/augmentation etc. Just saying I’m not sure we’ll need frontier models for 90%+ of consumer use cases I think the prevalence of benchmarking in the launch of new AI models is in agreement with this. This week I tried Grok 4.5 and Fable for some coding experiments and you need to really spend time to pick up the differences. So we use benchmarks to point out what’s not so obvious Some of us will remember when computers were all measured in megahertz and megabytes, and the PC industry compared itself that way. Over time, that gave way to design, power efficiency, etc. Today we’re benchmarking and calculating cost per token and so on. It’s about to evolve, I think
@interconnectsai ·
Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others The top end of the market was quiet, but "industry-scale tinkering" just got very loud. We're seeing a massive shift: specialized, cheap open models are now the crucial tools complementing closed agents. Analysis & breakdowns of: - @NVIDIAAI Nemotron-3-Super-120B: 1M context + NVFP4. - @cohere Transcribe - @SarvamAI -105B - @MistralAI -Small-4-119B Massive coverage of new artifacts from: - Coding & Logic: @zeddotdev @Meituan_LongCat & Goedel-LM - Multimodal/OCR: @Microsoft @YuanAI_Lab @BaiduAI & @xiaohongshu (RedNote) - RAG & Search: @trychroma @LightOnIO & @miromind_ai - Robotics/Agents: @Reka_AI @hcompany_ai & @ServiceNowRSRCH - Creative & Audio: @bfl_ml @hume_ai & GAIR - Data & Infra: @allen_ai @IBM @markov__ai @StepFun_ai By @natolambert & @xeophon https://t.co/NvFksswX9n
@neil_chilson ·
Very interesting! @DigitalFuture interview on China’s open source AI approach with Michael Kuiken of the U.S.-China Economic and Security Review Commission. The Commission, created by Congress, has a new report that finds that China’s open source / open weights approach “is accelerating innovation and adoption in ways the U.S. isn’t prepared to counter.” China's open source support is a key part of the deployment prong of their strategy. They really want to get AI out into the economy -- not just to deliver new functionality, but also to gather new data that will then improve their models. Meanwhile we're making it more difficult to deploy AI, both through increased red tape and fear mongering that is increasing skepticism.
@cyrilgupta ·
Jensen Huang shared a joint industry letter signed by NVIDIA and other leading tech companies on why open-weight AI models matter. Open models aren't just about transparency, they're about opportunity. For small business owners, they mean: • Lower AI costs • Faster innovation • Less dependence on a single vendor • The ability to build custom solutions without enterprise-sized budgets For SaaS founders and builders, open models unlock: • Faster product development • Greater flexibility to fine-tune models • Better control over performance, privacy, and deployment • More room to experiment without worrying about API pricing alone The next generation of successful businesses won't just use AI, they'll build unique products and workflows on top of it. Open models make that future accessible to everyone, not just the biggest tech companies. That's a win for innovation, competition, and entrepreneurship.
@interconnectsai ·
Latest open artifacts (#18): We analyzed 15+ new open models this week, uncovering some underrated gems and niche specialists. Featuring breakdown & analysis of: 🔹@LiquidAI's LFM 2.5 (1B) 🔹@arcee_ai's 400B MoE 🔹 New @Kimi_Moonshot updates 🔹 Specialized models from @NVIDIAAI, @GoogleDeepMind, @allen_ai, @AI21Labs, & @TIIuae 🔹 Plus: @Zai_org , @StepFun_ai and new @Alibaba_Cloud Qwen updates. By @natolambert and @xeophon https://t.co/EJXwoPgkwK
@suraj_sharma14 ·
Open source AI is having a moment. And it's not happening by accident. Here's why: 1/ It's now clear you can get close to frontier-level performance while keeping a transparent training pipeline. Today's Inkling launch from @thinkymachines is another strong example. 2/ More well-funded teams are building open-weight models than ever before. With models like Grok, Cursor, Muse Spark and others entering the space, we're heading toward a much more diverse ecosystem for coding and AI agents. 3/ Companies want more control over their data. Many are willing to trade a bit of frontier performance in exchange for privacy, ownership and not helping train a future competitor. 4/ Open source gives you flexibility. Bring your own model, evaluation stack, business context or harness. You're not locked into a single vendor. 5/ The conversation inside companies has shifted. It's no longer "How do we get everyone using AI?" It's "Why is our AI bill growing faster than the value we're getting?" 6/ Geopolitics matters too. More countries are looking at open-weight models as the best way to deploy frontier-level AI inside secure, controlled environments. We're moving from a world dominated by a few models... to one with real choice.
@CRSegerie ·
"Open source AI reduces power concentration" - it depends on what you mean by power. Open-weight models are genuinely useful: Universities can experiment freely. Self-hosting protects you from surveillance and political dependence on US providers. Per Epoch AI, the capability gap is only ~3 months on average. That's real access. But access to comparable capabilities ≠ power over the trajectory. By default, three US labs will decide what gets built next, what safety measures exist and what capabilities to pursue or hold back. Open-sourcing weights lets you use frontier AI. It doesn't give you a seat at the table where the frontier is shaped. I wouldn't be surprised if, at some point, they concentrate on a double-digit percentage of the economy. Open source reduces dependency. It doesn't reduce the concentration of the power that actually matters: who decides where this goes.
@_theshash ·
"The open-versus-closed debate in robotics isn't really about the models. It's about the data loop." — Samy Karim, founder of @cybernetic_lab Everyone imports the LLM frame: open weights vs closed weights, Llama vs GPT. Samy's point is that robotics isn't there yet. The training data barely exists. Teams are generating it right now, in simulation and through teleoperation, one task and one gripper at a time. So the real question isn't whether the model is open. It's whether the loop producing the data is open or closed. In robotics, that loop is already cracking open. Hugging Face's LeRobot is about a year old (it launched last summer) and it's already a hub: open hardware, open datasets, open model launches. @BitRobotNetwork put its Earth Rover hardware and the data it collects on there. Amanda, BitRobot's COO, runs what she calls an open robotics lab, pulling in contributors worldwide to collect data and evaluate models. On the eval side, platforms like RoboArena let you rank one model against another across environments, and both Nvidia and Physical Intelligence publish there. China has its own version, RoboChallenge. Samy's read: Chinese labs treat open-sourcing as a strategy, a way to build a moat and fundraise. So why does open have a shot here, when most frontier robotics work sits behind closed doors? Diversity. The labs want every kind of task, environment, and embodiment, and one warehouse is different enough from the next (different hardware, different grippers, different lighting) that no single company's data covers the space. That edge-case sprawl is exactly what a distributed network is built to collect. Then the part nobody mentions: you can't run these models out of the box. The post-training loop, fine-tuning the thing to actually work on a client's site, is where a deployment lives or dies. Open models make that loop workable for everyone, not just the lab that trained it. The closed loops are real. Tesla will have a great one for Optimus. China's data farms are running teleoperation at huge scale. But how much of that reaches anyone outside? So far, nothing. "It's the third time I've heard crypto is dead," Samy said about the broader gloom. "You can't kill what's already dead." Same logic on the data loop: it started about a year ago, and it isn't slowing down. Robotics Livestream EP3, watch it on YouTube below: https://t.co/kJmaWNeoIz
@pankajkumar_dev ·
Xiaomi MiMo-V2.5: Open-Source Model Focused on Efficiency and Long Context - Released under a clean MIT license, MiMo-V2.5-Pro is now tied with Moonshot’s Kimi K2.6 at the top of open-weights models (54 on the Artificial Analysis Intelligence Index) - Both Pro and base models support 1M token context, a major jump from the previous release - Leads open-weights on agent tasks (GDPval-AA 1578), ahead of DeepSeek V4 Pro and GLM-5.1 - Much more token-efficient: 92M output tokens vs 170M for Kimi K2.6, cutting inference cost nearly in half - Uses a 1T parameter MoE architecture (42B active) for strong performance with efficient compute - Around 75% non-hallucination rate, one of the better reliability ratios in open models - Day-one support via SGLang and vLLM, with optimizations for AWS and AMD hardware - Full FP8 needs 316GB VRAM, but lower-bit quants are expected to run on high-end local setups Feels like open-source models are moving toward long context, better efficiency, and real agent workflows, making self-hosting more practical over time.
@WesRoth ·
Anthropic is now the only major frontier AI lab that has not joined the industry’s defense of open-weight models. Nvidia, Microsoft, Meta, OpenAI, Google, SpaceX, Mistral, Hugging Face, and dozens of other organizations have backed the “Open Weights and American AI Leadership” letter. Anthropic’s absence is drawing criticism from investors, policymakers, and rival AI employees who argue that the company is using safety concerns to protect its closed-model business. But Anthropic’s position is not new. CEO Dario Amodei has repeatedly warned that once frontier model weights are released, the developer can no longer revoke access, update safeguards, or prevent people from modifying the system for harmful purposes. Anthropic instead favors gated access, where selected organizations can use powerful models while the company retains control. Anthropic may genuinely believe open frontier models are dangerous. Its critics may also be right that closed access protects its business. Both things can be true. The question is no longer whether Anthropic supports open source in principle. It is whether any company should be trusted to decide when intelligence becomes too powerful for everyone else to control.
@ForwardFuture ·
“Open models don’t mean open AI.” @hstyagi Co-Founder @SentientAGI says: “You can build on open-source models, but the data, feedback, and iteration loop are still closed.” “The real gap isn’t just the model, it’s the entire development process.” “We want that whole evolution of software and agents to be open.” “Not just models competing, but the systems that build agents competing in the open.”
Watch video
@shawnchauhan1 ·
The open-source AI landscape is fracturing at exactly the wrong moment. Meta has slowed Llama releases. DeepSeek R2 is delayed. Qwen's team is losing people. Developers who built routing architectures around these models are now exposed. Nvidia is stepping into that vacuum with Nemotron - and unlike the labs, Nvidia profits whether the model wins or the hardware wins. The open-source ecosystem assumed competition among many providers. It may be consolidating toward one with a very clear hardware agenda.
@shawnchauhan1 ·
Open-source AI was supposed to democratize intelligence. Alibaba just reversed course on Qwen, its open-weight leader. A few months ago, releasing weights openly was a competitive signal - proof of confidence, a bid for developer loyalty. Now it is a liability. When compute is constrained and inference is expensive, giving away the model means giving away the margin. The open-source era of frontier AI may be shorter than anyone predicted.
@zacodil ·
Open source was built for human collaboration. AI agents break that model. Agents don't need collaborators. One agent can ship an entire project in an evening. And when multiple agents work on the same codebase it gets worse, not better - they generate code faster than anyone can review, merge conflicts multiply, context drifts. The coordination problem that open source solved for humans becomes harder with agents. Proprietary models are far better at coding than open source ones. Claude and GPT one-shot tasks that local open models can't handle after dozens of attempts. For serious programming, the gap is still massive. So what's left for open source? The role is shifting. Open source used to mean "how we build" - many hands, shared effort, collaborative development. Now it's becoming "how we verify" - transparency, audit, trust. The code might be written by one agent on a proprietary model. But the result needs to be open so humans can check what the agent actually built. Open source isn't dying. It's more essential than ever - just not for the reason you think. Development goes solo. Trust stays open.
@Awesome_O_AI ·
Open source AI just stopped playing catch-up. Moonshot AI's Kimi K3 looks like the first open model that genuinely feels like a frontier model. Here's why that's a big deal 🧵 1/ Most AI demos show: • write an email • summarize a PDF • answer questions Kimi K3 was tested on tasks that take 30–40 minutes of autonomous work. That's a completely different category of AI. 2/ It doesn't just generate code. It plans. Uses tools. Tests its own output. Takes browser screenshots. Finds bugs. Fixes them. Repeats until the result actually works. That's agentic AI. 3/ One prompt produced a complete liquid physics simulator. No Three.js. No external libraries. Adjustable gravity, viscosity, lighting, and even webcam hand tracking. From scratch. 4/ Another prompt connected to Blender through MCP and built an animated V8 engine automatically. Planning. Modeling. Animation. Lighting. Rendering. All without manual intervention. 5/ It also created: • a playable 3D piano • a third-person shooter game • an AI-generated financial presentation with charts and narration • music composition software • deep research reports with citations These aren't toy demos. They're end-to-end workflows. 6/ The specs are equally impressive. • 2.8T parameters • 1M token context • Native vision • API access • Open weights planned Built specifically for coding, reasoning, and long-horizon autonomous tasks. 7/ The benchmark claims are hard to ignore. According to the presented results, Kimi K3 is competitive with leading proprietary models across: • Software engineering • Agent benchmarks • Frontend coding • Visual reasoning While also aiming for significantly better cost efficiency. 8/ The biggest takeaway isn't one benchmark. It's the trend. Open-source models are no longer months behind. The gap between open and closed AI is shrinking faster than almost anyone expected. 9/ If this pace continues, choosing an AI model may soon come down to workflow, tooling, and ecosystem, not whether it's open or closed. That would be a massive shift for developers, startups, and the entire AI industry. The next year is going to be interesting.
Best Open-Source AI tweets
Xholic studies what works in your niche, drafts posts in your voice and schedules them for the hours your audience is online.
No card needed · Nothing to cancel
Browse all tweet collectionsLive demo: no sign-up or login needed · takes about 2 minutes
Keep exploring