01
Best tweets about Mistral AI
50 Best Tweets About Mistral AI (2026)
Find the best tweets about Mistral AI, including model releases, Le Chat, Codestral, open models, benchmarks, and developer workflows. Updated weekly.
Specific Mistral model and product analysis, benchmarks, deployment experience, and developer examples.
- Creators
- 40
- Updated
Best Mistral AI tweets
Top Mistral AI tweets from 40 creators
Ranked 01–50
-
- 02
@MistralAI ·
Introducing Voxtral Transcribe 2, next-gen speech-to-text models by @MistralAI. State-of-the-art transcription, speaker diarization, sub-200ms real-time latency. Details in 🧵
Watch video
- 03
@kimmonismus ·
Mistral AI released Voxtral TTS, a 3-billion-parameter text-to-speech model with open weights that the company says outperformed ElevenLabs Flash v2.5 in human preference tests roughly 63% of the time on standard voices and nearly 70% on voice customization. The model runs on about 3 GB of RAM, achieves 90-millisecond time-to-first-audio, supports nine languages, and can clone a voice from just five seconds of reference audio, including cross-lingual adaptation that preserves the speaker's accent.
- 04
@mistralvibe ·
Mistral Vibe 2.0 is now available on Le Chat Pro and Team plans. Build, maintain, and ship code faster with the terminal-native coding agent by @MistralAI. Here’s what’s new 🧵
Watch video
- 05
@_avichawla ·
An open-weight alternative to ElevenLabs! Voxtral is a TTS model by Mistral with: - just 4B params - 70ms latency for voice agents - voice cloning from 3s of audio - 9 languages + cross-lingual transfer - 68.4% win rate over ElevenLabs Flash v2.5 Open weights on Hugging Face.
Watch video
- 06
@jeffboudier ·
What a week for open audio models! 🕺 💃 I demo: 🗣️ Voxtral 4B TTS from @MistralAI 🎙️ Transcribe 2B from @cohere 🏭 and how to run a batch transcribe job in 1 line of CLI using @vanstriendaniel uv script links below
Watch video
- 07
@AlbertQJiang ·
https://t.co/TMFEJO05cE Excited to share what the formal team @MistralAI has been building for the last couple of months: an Apache 2 Lean code agent with 6B active parameters. Outperforms open models like Qwen3.5, GLM5, Kimi-K2.5, and very competitive against Claude 4.6. Use for free in Mistral-Vibe: ``` $ uv tool install mistral-vibe --upgrade $ vibe / leanstall ``` enjoy the lean mode via shift+tab!
- 08
@emollick ·
So we now have a pretty good picture of the state of the frontier AI model makers. US closed source models continue to lead. Google, OpenAI, and Anthropic stand well ahead of the pack, and may have signs of recursive self-improvement. xAI has fallen from frontier status for now (though promises to return shortly). Meta re-entered the space today with a not-quite-frontier closed source model, but an approach that suggests that they might be back in the race. All the other US players seem far behind. On the Chinese model front, Alibaba (Qwen), Moonshot (Kimi), MiniMax, Xiaomi (MiMo), Deepseek, and Z (GLM) all still appear to be very much in the race, though the best Chinese models are still 7-9+ months behind released US closed source models. For some of these players, especially Xiaomi and Alibaba, their commitment to open weights appear to be slipping. Outside of China, Mistral seems to have fallen from frontier status.
- 09
@alex_verem ·
🚨Holy shit. Mistral just killed the ElevenLabs moat. > Voxtral TTS is open-weight, > Clones any voice from 3 seconds of audio, > Runs in 9 languages, > Beats ElevenLabs Flash v2.5 with a 68.4% human preference win rate. ElevenLabs built a moat on proprietary weights and API lock-in. Mistral just put the weights on Hugging Face. The model captures not just the voice but the person. Accents, inflections, intonations, vocal fillers the "ums" and "ahs" that make a voice sound human instead of synthetic. From 3 seconds of reference audio. Zero fine-tuning. Zero shot. The numbers: → 68.4% win rate against ElevenLabs Flash v2.5 in zero-shot multilingual voice cloning → Beats ElevenLabs Flash v2.5 on every one of the 9 supported languages → Matches ElevenLabs v3 on emotional expressiveness and quality → 70ms model latency same time-to-first-audio as Flash v2.5 at higher quality → 4B parameters. Runs on 3GB RAM. Smartphone. Laptop. Edge devices. → 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic → Cross-lingual voice cloning French voice prompt generating English speech works out of the box The strategic bet Mistral is making: enterprises don't want to rent a voice. They want to own it. Open weights means the model ships to your infrastructure. Mistral never sees your data. For healthcare, finance, and government Mistral's core customers that's not a nice-to-have. It's a compliance requirement. ElevenLabs built a $1B+ valuation on being the only serious option for production-grade voice AI. Today that changes. The weights are on Hugging Face. The API is $0.016 per 1,000 characters. Any developer can clone a voice, run it locally, and ship a voice agent without sending a single audio byte to a third party. One honest caveat: ElevenLabs shipped v3 after this evaluation. Voxtral matches v3 on quality metrics but the 68.4% win rate was head-to-head against Flash v2.5. The race isn't over. It just got a lot more competitive. Voice AI just got its open-source moment.
- 10
@GuillaumeLample ·
Very excited to release Voxtral 2! Voxtral 2 comes with two powerful new models: Voxtral Realtime: a SOTA real-time transcription model released under an Apache 2 license, with latency configurable to sub-200 ms Voxtral Mini Transcribe 2: a SOTA transcription model with speaker diarization, word-level timestamps, and context biasing Voxtral 2 supports 13 languages. It’s available on the Mistral API and is one of the most cost-effective transcription APIs on the market.
- 11
- 12
@akshay_pachaar ·
Transformer and Mixture of Experts in LLMs, explained visually! Mixture of Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? ____ Find me → @akshay_pachaar Every day, I share tutorials and insights on ML, LLMs, and AI Engineering.
Watch video
- 13
- 14
@ihteshamali ·
Arthur Mensch confirmed on CNBC this week that OpenAI and Anthropic are calling Mistral asking for compute. The two companies racing to build the most powerful AI in the world are dependent on a European competitor they have been trying to outrun. Let that land for a second. The narrative you have been sold is that the US is running away with AI. That OpenAI and Anthropic have such a commanding lead that the rest of the world is playing catch-up. That narrative is missing something physical. The International Energy Agency projected this year that US data center electricity demand will double by 2030. The US grid is already at capacity. Europe has the power. Europe has the land. And right now the company fielding calls from the most funded AI labs on earth is a four-year-old French startup. Mensch said Mistral is prioritizing its own customers first. The surge in agentic usage is generating far more tokens than anyone planned for and existing capacity is spoken for. The US labs are in the queue. But they are in the queue. Most people following AI are watching benchmark scores and funding rounds. The actual constraint right now is electricity and physical space. And on that dimension, the scoreboard looks nothing like the one most people are reading. Watch the full podcast on YouTube at @CNBCi
Watch video
- 15
- 16
@kimmonismus ·
Mistral just dropped Voxtral TTS an open-weight text-to-speech model with ultra-low latency, emotional expressiveness, and support for 9 languages. It outperformed ElevenLabs v2.5 Flash in zero-shot voice cloning tests judged by native speakers.
Watch video
- 17
@SebJohnsonUK ·
BREAKING: Mistral AI has raised a WHOPPING $830m in debt to build Nvidia-powered AI centres! The first facility is being built near Paris and will start operations before the end of June. It will house 13,800 of Nvidia’s top-end GB300 AI chips. Arthur Mensch, co-founder and CEO, said: "Scaling our infrastructure in Europe is critical to empower our customers and to ensure AI innovation and autonomy remain at the heart of Europe,” The company 20x’ed revenue in the last year, surpassed $400m of ARR and is on track to surpass $1bn this year. It has benefitted massively from Europe's push to sovereignty and also plans to invest €1.2bn to build out new AI data centres in Sweden. The financing has come from a bunch of banks, including Bpifrance, BNP Paribas, HSBC and MUFG. Europe is BUILDING - LETS GO.
- 18
@ArtificialAnlys ·
Microsoft has released MAI-Transcribe-1: a speech transcription model achieving 3.0% on AA-WER (#4), and is fast at 69x real-time The model was developed by Microsoft AI (MAI)’s Superintelligence team and supports 25 languages including English, French, Arabic, Japanese, and Chinese. MAI-Transcribe-1 API is currently available in public preview via Azure Speech on Microsoft Foundry. On the Artificial Analysis Speech to Text (STT) leaderboard, MAI-Transcribe-1 achieves a 3.0% word error rate on AA-WER for speech transcription accuracy, positioning it 4th overall behind Mistral’s Voxtral Small (2.9% AA-WER), Google’s Gemini 3.1 Pro High (2.9% AA-WER) and ElevenLabs’ Scribe v2 (2.3% AA-WER). It also stands out as one of the faster high-accuracy transcription models available, processing audio at ~69x real-time. See more details below ⬇️
- 19
- 20
@TheTuringPost ·
.@MistralAI's new Voxtral TTS generates expressive, multilingual speech from just ~3 seconds of reference audio It solves one of the hardest problems in speech, separating what you say from how you sound ➡️ Voxtral factorizes speech into two parts: • semantic tokens → the content (words) • acoustic tokens → the voice (tone, prosody, style) and uses different models for different problems + Voxtral Codec that compresses speech into ultra-low bitrate tokens This gives: - High-quality voice cloning that works best with ~3–25 second - support for 9 languages - 68.4% win rate vs ElevenLabs Flash v2.5 in voice cloning Here's how it works:
- 21
@Sumanth_077 ·
Clone any voice with just a 3-second audio clip! Mistral just open sourced Voxtral TTS, a text-to-speech model that clones voices from 3 seconds of audio and runs on edge devices. Here's what makes it different. Most TTS models need cloud GPUs and long audio samples. Voxtral runs locally on your phone or laptop and generates voice clones from 3 seconds. The model is 4B parameters. Small enough to fit on a smartphone but accurate enough to capture accents, inflections, intonations, and even vocal fillers like "ums" and "ahs" from the reference voice. It beat ElevenLabs Flash v2.5 in human preference tests with a 68.4% win rate. Key features: • 9 languages supported: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic • Zero-shot cross-lingual voice adaptation: Feed it a French voice sample and English text, get English speech with a French accent • Real-time performance: 90ms time-to-first-audio, renders 10-second clips in 1.6 seconds • Built for live voice agents and real-time applications Voxtral TTS is part of Mistral's full audio stack alongside their transcription and real-time STT models. Link to the model weights and research paper in the comments!
- 22
- 23
@mistralvibe ·
Here's another💎 from our recent hackathon in Tokyo. Adam Lensenmayer and Leonard Lin built "Vibecheck", a mobile app that lets you control your Mistral Vibe agents from your phone. Includes tool call approval, talking to your agent, and notifications when it needs you.
Watch video
- 24
@AbdelStark ·
I just shipped "parler" (French: *to speak*): Multilingual voice intelligence built on Mistral Voxtral model, generating decision logs from French/English meetings. Cocorico a bit 🇫🇷, but unironically: this is the kind of French AI stack I want to see being built, deployed on the industrial level, and win more. Sovereign AI is such an important matter, not only for us all as individuals, but also strategically for Nation States, it will become more and more a major geopolitical issue. Need more counter balance to the US frontier labs. We cannot let intelligence and knowledge be controlled by a handful of players, especially not concentrated in few countries. It leverages 2 pure French models from @MistralAI: - voxtral-small-latest for multilingual speech-to-text with timestamps - mistral-medium-latest for structured JSON extraction of decisions, commitments, open questions, and rejected options https://t.co/EVYMULKrmB
Watch video
- 25
- 26
@ArtificialAnlys ·
Mistral has released Mistral Small 4, an open weights model with hybrid reasoning and image input, scoring 27 on the Artificial Analysis Intelligence Index @MistralAI's Small 4 is a 119B mixture-of-experts model with 6.5B active parameters per token, supporting both reasoning and non-reasoning modes. In reasoning mode, Mistral Small 4 scores 27 on the Artificial Analysis Intelligence Index, a 12-point improvement from Small 3.2 (15) and now among the most intelligent models Mistral has released, surpassing Mistral Large 3 (23) and matching the proprietary Magistral Medium 1.2 (27). However, it lags open weights peers with similar total parameter counts such as gpt-oss-120B (high, 33), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, 36), and Qwen3.5 122B A10B (Reasoning, 42). Key takeaways: ➤ Reasoning and non-reasoning modes in a single model: Mistral Small 4 supports configurable hybrid reasoning with reasoning and non-reasoning modes, rather than the separate reasoning variants Mistral has released previously with their Magistral models. In reasoning mode, the model scores 27 on the Artificial Analysis Intelligence Index. In non-reasoning mode, the model scores 19, a 4-point improvement from its predecessor Mistral Small 3.2 (15) ➤ More token efficient than peers of similar size: At ~52M output tokens, Mistral Small 4 (Reasoning) uses fewer tokens to run the Artificial Analysis Intelligence Index compared to reasoning models such as gpt-oss-120B (high, ~78M), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, ~110M), and Qwen3.5 122B A10B (Reasoning, ~91M). In non-reasoning mode, the model uses ~4M output tokens ➤ Native support for image input: Mistral Small 4 is a multimodal model, accepting image input as well as text. On our multimodal evaluation, MMMU-Pro, Mistral Small 4 (Reasoning) scores 57%, ahead of Mistral Large 3 (56%) but behind Qwen3.5 122B A10B (Reasoning, 75%). Neither gpt-oss-120B nor NVIDIA Nemotron 3 Super 120B A12B support image input. All models support text output only ➤ Improvement in real-world agentic tasks: Mistral Small 4 scores an Elo of 871 on GDPval-AA, our evaluation based on OpenAI's GDPval dataset that tests models on real-world tasks across 44 occupations and 9 major industries, with models producing deliverables such as documents, spreadsheets, and diagrams in an agentic loop. This is more than double the Elo of Small 3.2 (339) and close to Mistral Large 3 (880), but behind gpt-oss-120B (high, 962), NVIDIA Nemotron 3 Super 120B A12B (Reasoning, 1021), and Qwen3.5 122B A10B (Reasoning, 1130) ➤ Lower hallucination rate than peer models of similar size: Mistral Small 4 scores -30 on AA-Omniscience, our evaluation of knowledge reliability and hallucination, where scores range from -100 to 100 (higher is better) and a negative score indicates more incorrect than correct answers. Mistral Small 4 scores ahead of gpt-oss-120B (high, -50), Qwen3.5 122B A10B (Reasoning, -40), and NVIDIA Nemotron 3 Super 120B A12B (Reasoning, -42) Key model details: ➤ Context window: 256K tokens (up from 128K on Small 3.2) ➤ Pricing: $0.15/$0.6 per 1M input/output tokens ➤ Availability: Mistral first-party API only. At native FP8 precision, Mistral Small 4's 119B parameters require ~119GB to self-host the weights (more than the 80GB of HBM3 memory on a single NVIDIA H100) ➤ Modality: Image and text input with text output only ➤ Licensing: Apache 2.0 license
- 27
@chutes_ai ·
Model Spotlight: Voidai Umbra 24B roleplay model. LoRA fine-tune of Mistral Small 3.2. Merged weights, Apache-2.0. Not a general assistant with RP bolted on: this was trained from the ground up for character voice and scene momentum. ~166M tokens of roleplay and instruction data across 6 epochs. What it does well: - Character consistency across multi-turn scenes - Vivid narration without drowning in purple prose - Inherits solid instruction following from the Mistral 3.2 base - Merged LoRA = no adapter overhead at inference Where it falls short: - Repetition creeps in during long generations - Sometimes writes your character's dialogue for you - Multi-character scenes need explicit formatting prompts - Practical context is 8k–16k, not the full base range - SFT only: DPO refinement is planned but hasn't shipped If you're building RP apps or interactive fiction, do give it a spin. Try it 🔗 https://t.co/J0wncpNi8I
- 28
@aryanXmahajan ·
ran 1,500 companies through an AI enrichment pipeline came out the other side with a projected $7,500 bill per run changed which model tier we used for 3 of the calls next run: $147 same output quality. same data. same pipeline. the entire industry defaults to the most powerful model available for every single task this is driving to the mailbox in a Formula 1 car here's how to actually think about it: REASONING TASKS → Claude Sonnet/Opus — complex analysis, voice matching, multi-step reasoning. anything where the output quality directly affects revenue FORMATTING / EXTRACTION (you already have the data, you just need it structured) → GPT-4.1-mini — $0.40 per million input tokens. this is the go-to for classification, JSON extraction, reformatting. replaced GPT-4o entirely. cheaper and better. → Claude Haiku 4.5 — $1 per million tokens. same use case when you need Claude-quality instruction following without the reasoning overhead WEB RESEARCH & ENRICHMENT → Perplexity sonar — $0.006/call. factual lookups, news checks, basic research. most enrichment pipeline calls belong here, not at the tier above it → Perplexity sonar-pro — $0.025/call. cross-referencing multiple sources, financial synthesis, strategic extraction → Mistral Medium with web connector — ~$0.03/call. solid for company enrichment. strategic initiatives, financials, triggers. 97% cheaper than deep research for the same quality on standard enrichment tasks → sonar-deep-research — $0.75-$1.00/call. 125x more expensive than sonar. use it when you genuinely need 20+ search multi-hop browsing. almost never for enrichment pipelines. SPEED-FIRST INFERENCE → Groq — open source models (Llama, Mixtral, Gemma) at inference speeds closed APIs can't match. matters when pipeline response time is the bottleneck OPEN SOURCE → Llama 3, Mistral, Qwen, Deepseek R1 — self-hosted. when data privacy matters, when margins are thin at scale, or when you need to fine-tune on proprietary data. the gap between open and closed is closing faster than most people realize FRONTIER REASONING → GPT o3/o4 — OpenAI's reasoning ceiling. use when you need it. prohibitive at pipeline scale → Grok 3 — xAI's flagship. stronger than most people give it credit for on reasoning and real-time data. it's connected to X natively — underutilized for anything requiring social signal synthesis or market intelligence the rule is simple: match model capability to task complexity most pipeline steps don't need intelligence. they need reliable formatting at cheap compute. that's a $0.40/million tokens problem. not a $15/million tokens problem. the companies winning on AI infra in 2026 aren't the ones with the best models they're the ones who know which model to use for which job
- 29
- 30
- 31
@etnshow ·
BREAKING: French AI start-up Mistral has raised $830mn in debt financing to support the construction of Nvidia-powered data centres across Europe. CEO @arthurmensch says: "Scaling our infrastructure in Europe is critical to empower our customers and to ensure AI innovation and autonomy remain at the heart of Europe". The investment comes as the company undergoes rapid growth from companies and governments seeking “sovereign” European alternatives to US groups such as Microsoft, Amazon and Google.
- 32
@Sumanth_077 ·
Make your CLI 10x more powerful with custom subagents! Mistral Vibe is an open-source coding agent that runs in your terminal. Most terminal agents give you one assistant that does everything. With Vibe, you can build custom subagents for targeted tasks like deploy scripts, PR reviews, and test generation, then invoke them on demand. Here's how it works: Each subagent runs independently and handles its job without cluttering your main workflow. You might have one for PR reviews, another for testing, and another for deployment. The agent also has full access to your codebase. It scans your project structure and Git status automatically, so it understands context across multiple files, not just the one you're editing. Key capabilities: • Web search and fetching - query and fetch content from the web mid-session for documentation, or API references. • Multi-choice clarifications - when your instruction is ambiguous, Vibe prompts with options instead of guessing. You pick the right path before it acts. • Full skills support - slash commands for preconfigured workflows, custom subagents for specialized tasks, and extensibility through the Agent Skills standard. Vibe works in your terminal or connects to VS Code, JetBrains, and Zed through the Agent Client Protocol. Mistral Vibe is available through Le Chat Pro and Team plans. It's 100% open source Link to Mistral Vibe in the comments!
Watch video
- 33
@KanikaBK ·
THE LLM CHEAT-SHEET FOR HERMES + OPENCLAW AGENTS The community has flagged Claude Opus 4.6 underperforming lately while GLM 5.1 has exploded on the scene to claim frontier capabilities. A lot has changed since the last version. Here's what moved: GLM-5.1 just proved its frontier capabilities with #1 SWE-Pro globally, 8-hour autonomous execution, and cheaper than Opus on input. It earns a Tier 1 spot. Grok 4.20 enters Tier 2 with the lowest hallucination rate of any tested model, a native multi-agent API running up to 16 parallel agents, and a 2M context window. Gemini 3.1 Pro drops to Tier 3. The price and multimodal story is strong, but the new frontier bar left it behind on reasoning. Mistral Small 4 joins Tier 3. One model replacing three specialist pipelines (reasoning, vision, agentic coding) at $0.15/M input. Apache 2.0. Here's the full landscape: 18 models in 4 tiers. Tier 1 - Frontier Models - Claude Opus 4.6: #1 agentic terminal coding; watch for inconsistency reports - GPT-5.4: superhuman computer use, real planning. and introduced a $100/month plan - GLM-5.1: #1 SWE-Pro globally, 8-hour autonomous execution, MIT license Tier 2 - Execution - MiniMax M2.7: 97% skill adherence, built for agents. API only, not open weights - Kimi K2.5: long-horizon stability, agent swarm - Grok 4.20: lowest hallucination rate on the market, native multi-agent, 2M context - DeepSeek V3.2: frontier reasoning at 1/50th the cost Tier 3 - Balanced - Claude Sonnet 4.6: 98% of Opus at 1/5 the cost - GPT-5.4 mini: 93.4% tool-call reliability, runs on OAuth - Gemini 3.1 Pro: best multimodal value, native video+audio in one call - Qwen3.6 Plus: near-frontier coding, completely free via OpenRouter - Llama 4 Maverick: open-weight, self-host at zero marginal cost - Mistral Small 4: one model replacing three; reasoning, vision, agentic coding, Apache 2.0 Tier 4 - Local / $0 - Runs on 32GB RAM or less - Qwen3.5-9B: always-on subconscious loop, 16GB RAM, beats models 13x its size - Qwen3.5-27B: stronger instruction following, 32GB RAM - Gemma 4 31B: best local reasoning, Apache 2.0, commercial-ready - DeepSeek R1 distill: best chain-of-thought at $0 - GLM-4.5-Air: purpose-built for agent tool use and web browsing, not a trimmed general model
- 34
@MistralDevs ·
CHECK THIS OUT. Coding in augmented reality with @mistralvibe. Built at the Mistral AI Worldwide Hackathon in Paris by @Min0laa, @tr_mz_, and @glwesteneng. Shoutout to Voxtral and Devstral 2 models. Demo was run on a Meta Quest 3.
Watch video
- 35
- 36
@NoPriorsPod ·
.@MistralAI is an open-source company. That's very different from OpenAI, Google, and other frontier labs today. Why does CEO @arthurmensch think open source matters? He explains in this conversation. Links in the replies.
Watch video
- 37
@HaimantikaM ·
If you are building in/with AI, one topic for sure that you should have hands-on experience on is quantization (basically reducing model weight and that often leads of throughput and memory gain). I made a benchmarking video on running a stress test prompt on Mistral 7B Instruct model on a @digitalocean GPU droplet (H200). Watch it till the end to understand if quantization is the right infra choice or not. P.S. - If you are building AI infrastructure, let’s hop on a coffee chat, would love to learn from you. https://t.co/JRZknohw5b
- 38
@Prathkum ·
Timeline of open-weight models (roughly chronological): 2023: open weight models are toys. Llama 1/2 are fun to fine-tune but terrible to actually rely on. Everyone quietly still calls the closed-source API when the task matters. The gap is common knowledge and nobody argues about it. Early 2024: Mixtral shows up and it's the first time a lot of people believe this could be usable. Mid-late 2024: Llama 3 / 3.1 405B lands and for the first time an open weight model is being benchmarked seriously against GPT-4 class models. Late 2024 into 2025: DeepSeek is the actual shock. Not just a good open model but a good open model trained for a fraction of the assumed cost, from a lab nobody in the west was tracking closely. This breaks the mental model a lot of people had, which was: frontier = closed = US labs. All three assumptions take a hit at once. 2025: this is where it stops being "open models catching up" and starts being "open models forcing the roadmap." Qwen, DeepSeek, Llama successors, Mistral, Kimi, multiple labs now shipping models that trade blows with frontier closed models on real benchmarks, not cherry-picked ones. June 2026: Z releases GLM-5.2, and it ranks fourth overall and first among open models on Artificial Analysis's Intelligence Index, matching Anthropic's Opus 4.8 within about a point on agentic benchmarks at roughly a fifth of the cost. People went crazy and started comparing it with frontier models like Fable 5. GLM 5.2 might be 1-2% lesser scored on the benchmarks as compared to top frontier models but it is almost 6 times cheaper than those models. It makes sense for people to compromise 1-2% capability based on the value they get from the 6 times cheaper rates. July 2026: Then Moonshot didn't even let the dust settle. Kimi K3 landed at 2.8 trillion parameters, the largest open-source model in the world, and Vals AI placed it second overall, behind Fable 5 and ahead of GPT-5.6 Sol. Two open releases (GLM 5.2 and Kimi K3), one month apart, both trading blows with the actual frontier. Upcoming: Qwen 3.8, GLM 5.3, and more. Crazy time to be alive.
- 39
- 40
@shawnchauhan1 ·
The company that bet everything on OpenAI just signed a multibillion-dollar deal to run open-weight models instead. Microsoft and Mistral expanded their partnership this week, bringing Mistral's open-weight Medium 3.5 model into Azure, Microsoft Foundry, and Copilot Studio. The detail worth noticing: enterprises can now run frontier AI fully disconnected from the internet, a first for Microsoft's cloud. Thousands of GPUs are going into new European data centers specifically so regulated industries can keep their data, and their model weights, entirely under their own control. It lines up with what Microsoft's own leadership has said elsewhere: closed models can cost you twice, once in fees and again when you hand over what you put in every prompt. When your closest partner starts hedging toward open weights, that's not a trend. That's a verdict.
- 41
- 42
@shawnchauhan1 ·
The open-source AI coalition model is maturing into something more structured. Mistral joined NVIDIA's Nemotron Coalition as a founding member. Their first deliverable is a shared base model built on DGX Cloud. Mistral Small 4 ships simultaneously through Mistral's API, Hugging Face, and NVIDIA's tooling in one release. That is three distribution channels activated at once, backed by a €1.7 billion funding round and deepening hardware partnership. The gap between open and closed model ecosystems is closing faster than most closed-model providers want to acknowledge. The open-source strategy is no longer scrappy. It is now well-capitalized and infrastructure-backed.
- 43
@thedevchandra ·
most people have no idea what's coming Q1 is barely in the rearview and: → Anthropic just partnered with Google to build Nexus Data Centers in Texas (just in) → Anthropic accidentally leaked "Claude Mythos" -a new model tier above Opus. so powerful they flagged it as an unprecedented cybersecurity risk. they didn't even mean to tell us → Claude can now use your computer - assign it tasks from your phone, it executes on your desktop while you sleep → GPT-5.4 reads 1 million tokens at once - feed it your entire codebase or content library in one go → Mistral Forge lets companies train AI on their own data - proprietary data is now a moat you can actually build on → MCP just hit 97M installs and every major lab supports it - one standard now connects AI to all your tools. build once, works everywhere the models are getting smarter faster than anyone planned the acceleration is starting to feel brutal is AGI still 18 months away?
- 44
- 45
- 46
@dailydotdev ·
Competitors eating your lunch is a great motivator, apparently. Here's today's AI dev news. - @OpenAI internally flagged @AnthropicAI as "code red" after Claude Code reportedly became the default at most tech companies - @OpenAI shipped parallel subagents in Codex today, so your main context stays clean while specialized agents do the dirty work - @AnthropicAI catalogued their last two weeks: voice mode, a marketplace, eight enterprise plugins, and four Google integrations, which is a lot - @MistralAI's Leanstral beats models 60x its size on formal verification benchmarks, at $36 versus $549, which is hard to ignore The chart of which models devs actually reach for tells its own story right now.
- 47
@vesting_tv ·
Mistral (@MistralAI) built an 8B robot-navigation model that beats multi-sensor systems with one cheap RGB camera. It hit 76.6% on unseen routes. The next robotics winner may sell the software that makes every robot useful.
Watch video
- 48
- 49
@rohanpaul_ai ·
Samsung is discussing a €1B Mistral investment at €20B valuation. It would nearly double Mistral’s €11.7B valuation from its previous financing. In this deal Samsung brings memory chips, devices, manufacturing scale, and direct access to billions of users. Mistral brings models that governments can run with more control over data and infrastructure. Europe wants that control after foreign access restrictions exposed dependence on American AI providers. Microsoft’s separate infra agreement gives Mistral distribution, while Samsung could strengthen its hardware position. --- reuters .com/business/finance/samsung-talks-invest-mistral-20-billion-euro-valuation-ft-reports-2026-07-22/
- 50
Best Tweets by Topic