Find the best tweets about Meta Llama, including open-weight models, fine-tuning, benchmarks, local deployment, and developer use cases. Updated every two weeks.
Model-specific Llama research and engineering discussions, excluding references to the animal or unrelated products.
Original Xholic analysis
Across 32 posts from 29 creators, X conversation about Llama skews technical and supportive (90.6% SUPPORTIVE, 46.9% POSITIVE, 43.8% NEUTRAL). Local/on-device inference (28.1%), fine-tuning and adaptation (25%), and hardware benchmarking (25%) dominate, punctuated by a Meta-strategy debate and one major outlier post claiming 70B models run on a 4GB GPU (and Llama 3.1 405B on 8GB) via AirLLM's layer-by-layer loading (2014005554948047122).
46.9% of posts
All-time engagement
100% of posts
Published in 90 days
Conversation map
Running Llama models on consumer hardware without cloud APIs: llama.cpp and its origin story, AirLLM layer-by-layer loading, offline MacBook workflows (71 tok/s on flights), local-first positioning ('local models are the only saviors now').
28.1%
Adapting Llama models to specialized tasks and behaviors: QLoRA fine-tuning pipelines on tool-call datasets, LoRA SFT + DPO with distilled reasoning traces (Netflix RecSys), rebuilding RLHF pipelines from Llama weights, early Llama-1 fine-t
25%
Empirical throughput/latency comparisons on specific hardware: AMD EPYC host-node benchmarks, Intel Core Ultra/PyTorch edge inference, MLX vs llama.cpp decode speeds on Apple Silicon, GPU-vs-Mac for agentic parallel workloads, 70B stress
25%
Commentary on Meta's trajectory with Llama: Zuckerberg on why Llama 4 broke after Llama 3, Meta abandoning open models as a strategic fumble, distribution-vs-model-quality debate, nostalgia for the 'Zuck doing OSS' era.
18.8%
The broader open-weight ecosystem and where Llama stands in it: arXiv-wide usage analysis showing Qwen overtaking Llama, chronological timeline of open-weight model capability (Llama 1→3.1→DeepSeek→GLM/Kimi), 'toys to frontier' narrative.
12.5%
Running models at reduced precision: Q4_K_M / int4 / 2-bit weight-only quantization (KronQ post-training quantization on LLaMA-3-70B), quantization-stacking behavior across model sizes, torchAO integration for edge inference.
12.5%
Novel research applied to Llama models: evolution strategies as an RL alternative for fine-tuning, byteification retrofitting Llama 3 to byte-level models, data-centric training frameworks (DataFlex on LLaMA-Factory).
12.5%
Research showing orchestration/control matters more than scale: the Atomic Task Graph (ATG) thread where Llama-3.1-8B beats GPT-4+ReAct via DAG-based execution, model size vs. framework contrarian arguments.
6.3%
Tone and stance
Performance benchmark
Posts with media make up 78.1% of this collection. Their median all-time score is 6.74, compared with 1.85 for text-only posts.
Format mix
Consensus and debate
Shared view
Multiple independent posts converge on running Llama models without cloud APIs: the llama.cpp origin story (2063904209691685299), an 11-hour offline MacBook flight run of Llama 70B via llama.cpp at 71 tok/s with 60k context (2050462114167611869), AirLLM's layer-by-layer loading for 70B on 4GB VRAM (2014005554948047122), and 70B Q4_K_M stress tests on an M5 Max 128GB at 48K context (2038808152620052972). Local-inferer
Shared view
Posts converge on low-cost specialization of open-weight Llama models: a QLoRA fine-tune of Llama-3.1-Nemotron-Nano-8B for schema-valid JSON tool calls completed in ~3 hours on one L40S for under $10 (2072963351144009730), and a Netflix paper's LoRA SFT + DPO on Llama 3.1 3B/8B students beat its production RecSys by ~5% on inverse propensity score (2040879588822380554).
Open debate
One post argues GPUs are non-negotiable for parallel/agentic workflows, citing Llama 3.1 70B BF16 on 8x RTX 3090s serving 50+ concurrent requests in 2m29s and stating 'You can't do this on a Mac. Not yet at least' (2040627991156343192), while local-first posts tout single-Mac Llama workflows (2050462114167611869). The tension is between throughput-heavy batch serving and single-user offline use.
Open debate
Posts diverge on Meta's position: nostalgia for the 'Zuck doing OSS' era and lament that Llama 4 changed it (2072415565994455448), a post calling Meta's retreat from open models a strategic fumble given enterprise demand for cheaper local alternatives (2068071583240061010), versus a pod-based take that Meta's distribution (~3B users) is the real moat even if 'Llama 4? Open, flexible...but not leading' (2043838420896
What performs
The AirLLM announcement (2014005554948047122) scored 4954.32 all-time, ~1065x the median all-time score of 4.65 — far beyond the next outlier, the Atomic Task Graph thread (2076465883938009457, 195.58, ~42x median) claiming Llama-3.1-8B-Instruct beats GPT-4+ReAct on ALFWorld and WebShop via DAG-based execution.
Median engagement across the 32 posts is 30 likes, 3628 views, 3 replies, and 3 reposts (median all-time score 4.65). Posts with media (78.1% of the set) have a median all-time score of 6.743 versus 1.851 for text-only posts.
Median post age is 208.4 days, spanning 2023-03-02 to 2026-10-18; only 4 posts (12.5%) fall within the last 90 days, so the dataset mixes early Llama-era posts (1631778090748657664, 1631140243222167554) with recent ones.
Statistical standouts
Creator landscape
The five most represented creators account for 25% of the selected posts.
1. BURKOV
@burkov
2 posts
2. Hugging Models
@HuggingModels
2 posts
3. simp 4 satoshi
@iamgingertrash
2 posts
4. abdel
@AbdelStark
1 post
5. Alex the Engineer
@AlexEngineerAI
1 post
6. AWS AI
@AWSAI
1 post
32 posts come from 29 creators, and the top five voices account for only 25% of placements — no single account dominates the Llama conversation.
burkov posted 2 of the top research-methods posts with a median all-time score of 8.08: evolution strategies as an RL-free fine-tuning alternative for Qwen and Llama models (2041792519030128782) and KronQ 2-bit post-training quantization on LLaMA-3-70B (2080443570784190689).
Hugging Models (median all-time score 0.99) and iamgingertrash (0.58) each posted twice with below-median engagement; iamgingertrash's posts are historical March 2023 fine-tuning notes (1631778090748657664, 1631140243222167554), while Hugging Models' are model-release announcements (2040528091295203716, 2040535641642213557).
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 32-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 Llama tweets
Ranked 01–32
@LiorOnAI ·
You can now run 70B LLMs on a 4GB GPU. AirLLM just made massive models usable on low-memory hardware. 𝗪𝗵𝗮𝘁 𝗷𝘂𝘀𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱 AirLLM released memory-optimized inference for large language models. It runs 70B models on 4GB VRAM. It can even run 405B Llama 3.1 on 8GB VRAM. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 AirLLM loads models one layer at a time. Instead of loading everything: → Load a layer → Run computation → Free memory → Load the next layer This keeps GPU memory usage extremely low. 𝗞𝗲𝘆 𝗱𝗲𝘁𝗮𝗶𝗹𝘀 • No quantization required by default • Optional 4-bit or 8-bit weight compression • Same API as Hugging Face Transformers • Supports CPU and GPU inference • Works on Linux and macOS Apple Silicon 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 • Run Llama, Qwen, Mistral, Mixtral locally • Test large models without cloud GPUs • Prototype agents on cheap hardware
@IntuitMachine ·
The One Change That Lets Small Models Outperform Their Size 1/ Everyone knows you need a 70B model to beat GPT-4 on complex agent tasks. We did it with 8B—by changing one thing that has nothing to do with the model. A thread on why your agent's biggest problem isn't the LLM. 🧵 2/ The standard approach: feed the LLM a growing text history, ask it to pick the next action, repeat. This works... until it doesn't. Errors propagate. Context bloats. Hallucinations spike. And when something breaks, you replan everything. 3/ Here's the kicker: The problem isn't your model's intelligence. It's that you're asking it to hold plan structure + execution state + I/O dependencies all inside a linear text stream. That's like running an OS without a process table. 4/ Enter: Atomic Task Graph (ATG) Instead of a text trajectory, you build an explicit DAG. Each node = one tool call. Edges = data dependencies. The LLM still does the thinking—but now the graph holds the structure. 5/ Three moves make this work: ✅ Interface-preserving recursion: Break tasks into subtasks while keeping I/O contracts clean ✅ Dependency-aware execution: Run independent branches in parallel; catch bad plans before running them ✅ Minimal repair: When something fails, fix only the broken subgraph—leave the rest frozen 6/ Result? Llama-3.1-8B-Instruct beats GPT-4+ReAct on ALFWorld (household tasks) and WebShop (shopping). Not with fine-tuning. Not with more data. Just by swapping the execution substrate from text → graph. 7/ Why does this work? Context narrowing: Each node sees only its local inputs—no bloated history Pre-execution validation: The graph lets you "think" before acting Localized failure: Repair 10% of the graph instead of replanning 100% 8/ The contrarian insight: Control framework > model size (in the 7–70B range). You're not squeezing more juice from the same fruit. You're giving the model a better glass to pour into. 9/ Practical translation: • 20–40% step reduction (parallelism) • 70%+ hallucination drop (narrower context) • 3× faster recovery (minimal repair) • Training-free, plug into existing tool APIs This isn't research theater. It's production-ready architecture. 10/ The bigger implication: If you can beat GPT-4 by changing the substrate instead of the model, what else have we been over-parameterizing? Retrieval pipelines? Code generation? Multimodal workflows? The graph wins again. 11/ [Final + CTA] TL;DR: Stop storing your agent's plan in text. Start storing it in a DAG. Small models suddenly look a lot smarter.
@TheAhmadOsman ·
People ask why I keep insisting on GPUs and not Mac Studios/Mac minis for parallel & Agentic Workflows (multi-agents) This is why: - Llama 3.1 70B BF16 (~140GB w/o Context) - on 8x RTX 3090s - Synthetic data generation with - 50+ concurrent requests - Batch inference - Sustained throughput Not only that: > ~2k context per request (prompt) > ~1.8k tokens in output > 2 mins 29 secs for 50 responses This is GPU territory. You can’t do this on a Mac. Not yet at least.
@natolambert ·
Over the weekend I had Codex parse 500K arXiv AI/ML papers since ChatGPT to understand which open models are used for research. In 2024, ~30% of papers mentioned an American open model and only 10% a Chinese model. Today, ~40% of papers mention a Chinese (open) LLM, and only 25-30% an American one. Chinese models are the default for research. Chinese mentions are still growing while American open models are stagnating. When looking at this data it's important to remember that papers substantially lag model releases, as research takes a long time. Qwen's steady growth is reflective of this, but so is Llama's lasting power. Some more observations: 1. Qwen has been steadily growing, and today 1/3 of papers which mention any LLM mention qwen. OpenAI's closed models are the highest overall, at ~37%. 2. Llama peaked around April of 2025 at 30% of papers which mention any LLM (including ChatGPT etc). Llama 4 was released at about the same time, and Llama has been declining since. 3. Gemini and Claude are less common than the leading open models, mentioned in 10-15% of papers puts them behind all of Qwen, Llama, and DeepSeek. Open models should be and are the foundations of open research. The % of papers mentioning any LLM have been steadily climbing since 2023. | Year | January | April | July | October | | 2023 | 10.43% | 15.39% | 18.69% | 32.18% | | 2024 | 29.70% | 33.93% | 35.70% | 44.25% | | 2025 | 39.23% | 45.28% | 44.94% | 53.52% | | 2026 | 55.49% | 57.26% | 53.14% | TBD Now over 50% of AI papers, from 10% in 2023. Other notes: - Gemma and Mistral hover around 5-10%. - Our beloved fully-open Olmo models have been ~1% since the first release in Jan. 2024. - DeepSeek has a clear jump after R1 in Jan. 2025 - Data derived from the most popular ML arXiv categories: cs. AI, cs. CL, cs. CV, cs. LG, stat. ML Just like our downloads and derivative model data, this is updated daily on the Interconnects Open Model Dashboard.
@rohanpaul_ai ·
Mark Zuckerberg talks about why the Llama run broke after Llama 3. says Llama 3 was a very good model and almost at the frontier, then Llama 4 fell off the path they needed. His read is that he staffed it like Instagram feed and ads, with hundreds or thousands of people in parallel, when a frontier model needs a small tight science team. "When you're an entrepreneur, you're building something, I think you kind of get tested in those moments because inevitably not everything is going to go well. And the things that kind of define the trajectory are, okay, if something doesn't go the way that you want, how do you basically figure out how to move forward?" ---- From "The Next Big Thing" YouTube channel, (link in comment)
Watch video
@TheTuringPost ·
An interesting paper was just published by @allen_ai in Nature on ditching subword tokenization without burning millions of dollars in compute: Retrofitting language models to operate over bytes. The classic headache with byte-level models has always been speed and the sheer cost of training them from scratch. Instead of starting from zero, the authors introduce "byteification" – a neat two-stage distillation trick to retrofit existing pretrained models (like Olmo, Llama 3, and Qwen) into byte-level systems using under 1% of the original pretraining budget. The key architectural insight addresses a subtle flaw in earlier latent tokenizer setups. Subword tokenizers look ahead at future bytes when drawing boundaries, whereas standard byte-level patch predictors were strictly causal. By adding just 1 byte of lookahead during prefill and predicting fused boundary tokens during decode, they bridge that expressivity gap. The resulting models (Bolmo, Blama, Bwen) run at practical inference speeds, crush character-level reasoning tasks where standard tokenizers trip up, and you can even transfer instruction tuning across using simple task arithmetic. Definitely worth a close read if you work on token-free architectures.
@OpenGradient ·
OpenGradient Model Highlight: Dobby Mini Leashed Dobby Mini Leashed by @SentientAGI is a fine-tuned Llama 3.1 8B model exploring behavioral consistency and stable interaction patterns in open-source AI systems. 🧵👇🏻
@heyrimsha ·
A software engineer in Sofia, Bulgaria wrote 4,000 lines of C++ in March 2023 that made it possible to run Meta's leaked Llama model on a MacBook without a GPU. Within a week every AI engineer on Earth was running his code. 3 years later the project has 115,000 GitHub stars with powers most of the local AI ecosystem and just joined Hugging Face. He had never worked at a major AI lab in his life. His name is Georgi Gerganov and most people just call him ggerganov. Here is the story because almost nobody outside the open-source AI world knows what one engineer in Bulgaria has built. Georgi lives in Sofia. He works from a home office in a country that has no frontier AI lab, no NVIDIA partnership, no Silicon Valley. He had been writing low-level C and C++ for years before LLMs became the center of the universe. He was known in niche corners of the internet for an odd project called kbd-audio, a program that could figure out what someone typed by listening to the acoustic signature of their keyboard. In September 2022 he started working on something called GGML. It was a tensor library written in pure C with no dependencies. The inspiration was Fabrice Bellard's LibNC. The goal was simple. Run machine learning models on regular hardware with no Python, no PyTorch, no CUDA, no cloud. The first real test was whisper.cpp, his port of OpenAI's Whisper speech recognition model. It ran on a laptop. It ran on a phone. It needed nothing but a C compiler. The open-source community started noticing. Then in late February 2023 Meta released Llama, the first serious open-weight large language model. Within a week the weights leaked on 4chan and started spreading across the internet. The catch was that almost nobody could run it. You needed expensive GPUs and a Python environment. The most powerful open model in the world was effectively locked behind hardware most developers did not have. On March 10, 2023, Georgi pushed the first commit to a new repository called llama.cpp. It was an implementation of Llama inference in pure C and C++ with zero dependencies. It ran on CPU. It ran on a MacBook Air. It ran fast. The repository exploded. Within days it had thousands of stars. Within weeks it became the default way to run open-source LLMs on any computer. Quantization support came next, letting you run a 7 billion parameter model in 4GB of RAM. Then a 13 billion parameter model on a phone. Then a 70 billion parameter model on a single high-end consumer card. Today llama.cpp is the inference backbone of the local AI movement. Ollama runs on it. LM Studio runs on it. GPT4All runs on it. Most "run AI on your laptop" tutorials you have ever seen are running Georgi's code underneath. The repository has 115,000 GitHub stars. The community has added Vulkan, OpenCL, Metal, CUDA, and RPC-distributed inference. The K-quant compression methods alone changed how the entire ecosystem thinks about model size. In 2023 Georgi founded ggml AI in Sofia with pre-seed funding from Nat Friedman, the former GitHub CEO, and Daniel Gross. He stayed in Bulgaria. He kept the company small. He kept the code MIT licensed and free forever. In February 2026 Georgi and his core team, including Xuan-Son Nguyen and Aleksander Grygier, joined Hugging Face full-time. The deal was framed as a partnership to scale local AI as a serious alternative to cloud inference. The code stays open. Georgi still answers GitHub issues personally. He still posts releases without press releases. His personal website is a flat page. He still plays semi-professional basketball for a club in his hometown in Bulgaria on the side. A guy in Sofia who had never worked at OpenAI, Google, or Meta wrote the code that runs most local AI on the planet. He did it in a few weeks.
@eric_seufert ·
LLMs are increasingly being used in RecSys for personalization and ranking tasks, where semantic and contextual knowledge can be brought to bear to rank pieces of candidate content using sequences of a user's behavioral history. Netflix has a new paper out that explains how they've utilized LLMs to predict which artwork is most likely to resonate with a user when recommending a title to them. In the paper, the authors query a large LLM (Qwen 32B) with user-title interaction snapshots constructed from users' recent engagement histories: - user watch history up to that point - a new title that the user engaged with - the set of candidate artwork available for that title - the ground truth artwork the user engaged with In these queries, the prompt requests that the LLM provide a post-hoc justification for why the ground-truth engaged artwork might have outperformed (resonated more than) the others. With this data, the authors LoRA fine-tune a smaller LLM (LLaMA 3.1 3B and 8B variants) with the synthetic reasoning traces generated by the teacher model, using SFT, SFT + Reasoning, and DPO as fine-tuning regimes. They find that the 8B student model fine-tuned with SFT + Reasoning outperformed their production RecSys infrastructure by ~5% on inverse propensity score, which corrects for selection bias given differing sizes of candidate artwork sets across titles. What's fascinating about the use of LLMs for RecSys is how effectively open-weight models can be adapted to production RecSys tasks. The authors specifically avoided RL, opting instead for LoRA SFT and DPO, which is vastly cheaper to implement given the training data sizes used in the paper (10k and 110k samples). Paper linked below.
@burkov ·
For the past few years, the standard recipe for finetuning LLMs on tasks like math reasoning has been reinforcement learning (RL): you let the model generate answers, score them, and use the scores to nudge the model's parameters via gradients. RL has known weaknesses here—it struggles when rewards only arrive at the end of long answers, it often "hacks" the reward by finding degenerate shortcuts, and two runs with identical settings can end up with very different final performance. This paper shows that an old and much simpler family of methods, called evolution strategies, works well on models with billions of parameters, which most researchers had assumed was impossible. The method is straightforward: take the model, make thirty copies with small random noise added to every parameter, score each copy on the task, then shift the original parameters slightly toward the copies that scored higher. No gradients, no backpropagation, no value networks, no penalty terms to tune. Using this approach, the authors finetune models from the Qwen and Llama families on a symbolic arithmetic puzzle, several math benchmarks, and Sudoku, and match or beat welltuned RL baselines while keeping the same hyperparameters across every experiment. Read with an AI tutor: https://t.co/tqNaoPETaY PDF: https://t.co/cmPpsx3R1q
@TimJayas ·
JUST RAN LLAMA 70B LOCALLY ON A MACBOOK FOR 11 HOURS ON A FLIGHT WITH ZERO WIFI > No cloud APIs > No Anthropic / OpenAI servers > Just llama.cpp @ 71 tokens/sec > 60k context, 48.6 GiB memory used > Battery budget: 3h21m, checkpointed every 12 tasks no wifi. no API cost. cleared his entire client queue before landing. seeing the increase in AI subscriptions. local models are the only saviors now.
Watch video
@no_stp_on_snek ·
stress testing Llama-3.1-70B Q4_K_M on M5 Max 128GB. early results: turbo3 prefill is FASTER than q8_0 (baseline) at 32K context (80.8 vs 75.2 t/s). less KV bandwidth wins when the cache gets big enough. decode flat. PPL healthy across all configs ... no catastrophic failure yet. turbo4/turbo4 +6.3%, turbo3/turbo3 +11.4%. larger models absorb quantization stacking that kills 7B models. 48K context confirmed working. 70B on a laptop. still pushing for more. HUGE
@forgebitz ·
remember those 2 months everyone thought zuck was kinda cool and doing oss ai models wilde how llama 4 changed all of that
@sickdotdev ·
A developer reportedly ran Llama 3.3 70B locally on a MacBook Pro M4 during an 11-hour transatlantic flight, completing client work entirely offline without internet access. Using llama.cpp, the setup achieved about 71 tokens per second with roughly 60,000 tokens of context while using nearly 48.6 GiB of memory. Tasks were queued in a JSONL workflow, where the model processed one job at a time on localhost, saved results automatically, and created checkpoints every 12 tasks so work could resume smoothly if interrupted.
Watch video
@HuggingPapers ·
DataFlex A unified data-centric training framework built on LLaMA-Factory, supporting dynamic sample selection, domain mixture adjustment, and sample reweighting with full DeepSpeed ZeRO-3 compatibility.
@AbdelStark ·
Ok this is insane. Open source models and recipes for sovereign specific agentic workflows are becoming extremely accessible. I did a QLoRA fine tuning on nvidia/Llama-3.1-Nemotron-Nano-8B-v1 base model, to emit exactly one schema-valid JSON tool call per request. It took only about 3 hours on a single GPU (@nvidia L40S), and costed less than 10$. Summary of the end to end pipeline, ran remotely on the GPU using @modal infrastructure: 60,000 Salesforce xlam rows → 28,461 valid single-call → 26,735 deduplicated → 15k/1k/1k splits → base eval → QLoRA training (1,876 steps, 2 epochs, peak 26.3 GiB) → adapter eval → comparison with base You can see a significant jump in accuracy of the adapter over base model. Adapter checkpoints and datasets are published on @huggingface. The era of specialized models and agentic workflows will explode soon. Open source AI for the win. Sovereign AI for the win. It's happening.
@AlexEngineerAI ·
Stop using massive models for simple logic. Run Llama 4 Scout for your basic data parsing and routing. It is faster, keeps your data local, and costs $0. Save the heavy reasoning for your complex features.
@burkov ·
This joint work of @USC and @Yale scientists develops KronQ, a novel post-training quantization framework that achieves state-of-the-art 2-bit weight-only quantization on LLaMA-3-70B by incorporating gradient covariance through a Kronecker-factored Hessian, significantly outperforming existing methods that fail to converge. Read with an AI tutor: https://t.co/7u979gzBve
@PyTorch ·
PyTorch 2.10 is now optimized for @Intel Core Ultra Series 3 processors to bring high-performance AI to the PC and edge. This release leverages the new Xe3 architecture and Arc B-series GPUs to deliver up to 120 XMX TOPs. Native TorchAO integration enables seamless int4 weight-only quantization, allowing a Llama 3.2 3B Instruct model to run locally with a first token latency of 242.29 ms and a 2+ token latency of 27.24 ms. For edge scenarios, developers can achieve 1.4x to 1.7x faster training for vision models via Anomalib, with WinClip seeing gains up to 2.5x compared to previous generations. Read our latest blog from the Intel PyTorch and Client AI SW teams for the full technical deep dive and benchmarks: https://t.co/pEGpmwtROi #PyTorch #Intel #AIPC #EdgeAI #OpenSourceAI
@SaiyamPathak ·
Yesterday in the video I took - the qwen32B which was a dense 32B and all active parameters for every token whereas for the MLX version it was A3B - active 3B. this morning I ran some tests again: - Qwen3.5 (NVFP4, MLX): 23.11 tok/s decode - Nemotron (GGUF, llama.cpp): 21.70 tok/s decode So not that much of a difference - maybe llamacpp already good for M1 type - I need to test it across all architectures though. Interestingly I tested the LMstudio as well that had the MLX support for quite some time now. results are pretty interesting and I tried kind of similar models although its not apples to apples. LM Studio - MLX - Qwen3.5-35B-A3B-4bit - 33.82 tok/s Ollama - MLX - qwen3.5:35b-a3b-coding-nvfp4 - 23.11 tok/s the 4-bit quantization is different here for both models. Can you share your benchmarks? anything anyone tested.
@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.
@AWSAI ·
🧬 Strands Agents is enabling sophisticated multi-agent architectures 🤖⚡ Our latest blog showcases a powerful video processing workflow using Meta's Llama 4 on #AmazonBedrock. #AWS 👉 https://t.co/F7PS2gShSf
@Bozos ·
favorite detail in meta's llama 3 paper is that training ran 1-2% slower around midday because the gpus got hot. sixteen thousand h100s and it still had a rough afternoon
@pritopian ·
Meta abandoning open models feels like a big fumble. Enterprises are placing limits on token use, and are looking for cheaper and local alternatives. Meta was quite ahead at some point with Llama! They were well positioned to become the default foundation for enterprise AI. Missed opportunity.
@HuggingModels ·
Meet Dolphin3.0-Llama3.1-8B. This isn't just another language model. It's a specialized 8B parameter model fine-tuned for reasoning, math, and function calling. Think of it as a compact powerhouse built for complex tasks.
@boyuan_chen ·
DataFlex is a useful paper for one reason: it turns data-centric training from a pile of isolated repos into something you can actually compare and plug into an existing LLM pipeline. Built on LLaMA-Factory, it unifies 3 knobs in one framework: data selection, domain mixing, and sample reweighting.
@Signal_65 ·
What role does the host CPU play in AI inference performance? @Signal_65's latest research answers that question. https://t.co/rKI4hrsoic We benchmarked @AMD EPYC host nodes against competitive options across 7 distinct AI models, including GPT-OSS-120B, Llama-4-Scout, and DeepSeek-R1. The data shows clear advantages for AMD EPYC in throughput, time to first token, and inter-token latency. As AI workloads scale, the efficiency of the host node becomes a critical factor in overall system performance and TCO. Read the full report and watch our expert interview on https://t.co/puT6dxKVTC!
@HuggingModels ·
Meet Meta-Llama-3.1-8B-Instruct-GGUF: a powerful, efficient text generation model that's been downloaded over 240k times. It's the community's go-to for multilingual AI tasks, and it's ready to run on your local machine right now.
@RockportAI ·
If you're building AI for healthcare, stop. Read this first. A new RCT just published in Nature Medicine tested 1,298 real patients using GPT-4o, Llama 3, and Command R Plus for medical questions. The group with zero AI outperformed every AI group — by 76%. Not because the AI didn't know medicine. It did. The failure was the interface between human uncertainty and AI confidence. 15-minute read → 4:37 listen. By the research team for Nature Medicine.
Watch video
@iamgingertrash ·
We’re fine tuning llama right now, and first thoughts 1. They’ve left out the ‘backward’ function so we have to implement training ourselves 2. Context window it’s been trained on is 2048 so 1/2 the context that Davinci can take and 1/16 the context the new GPT-4 can take
@iamgingertrash ·
Got llama weights. Gonna try to rebuild RLHF pipeline via @huggingface datasets for instruct models Intuition is telling me this could come close to instruct gpt performance
@TheSixFiveMedia ·
Meta’s AI strategy isn't about building the best model, it's about building the most accessible one. @danielnewmanuv & @PatrickMoorhead break it down on Ep. 300 of The Six Five Pod, distribution is the real moat with ~3B users across Meta’s platforms. Muse Spark signals the next phase. Llama 4? Open, flexible…but not leading. Is Meta quietly winning AI by owning the audience?
Watch videoBest Llama 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