X Archive

Don't sleep on this model everyone, this is SOTA for <= 16GB memory for inference (which is a LOT of people, old M series macbooks, gaming laptops and such)

We’re open-sourcing Pro and Flash, MiMo-V2.6-Distill-Qwen-9B, the technical report, 7K+ RL task environments, an end-to-end RL framework and composable mini-harnesses. Reproduce, verify and build on the work.

too many open jev claims and reproductions which one works? which one can you run on your laptop? i created this tracker that categorizes 38 artefacts (github repos, hugging face models) and separates signal from noise huggingface.co/spaces/multimo…

RIP Terra. 2026-2026 💔

Today, we’re announcing Ternary Bonsai 2 27B. Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance. Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use. Ternary Bonsai 2 27B is available today under Apache 2.0.

I see pi-workflows viewers growing, so wanted to give a disclaimer: I did major refactors and broke some things in the last 2-3 minor versions I will post an update when it is back to full feature, and better than before

"Fair challenge", "Fair — that's on me", "That error is the whole bug", "Short answer: no. Not honestly" I've been hit by top 4 claudisms in the first 4 messages with DeepSeek V4.1 Flash I hate to be the bearer of bad news but this model is gonna SPREAD and it's gonna bring Claudish everywhere it goes 😭 Just when I thought the situation started to improve with OAI's newer generation of models...

Launching ourmodels.cc Front Page 🎉🎉🎉 If you are addicted to local/open-weight models like me, here is a new page to check daily, alongside Twitter and Hacker News Whenever someone here on Twitter posts about a model and it receives enough engagement, I scrape it and add to my knowledge graph There is so many things happening in open models specifically, that one needs an aggregator to keep track of everything! Who are some new people who got in open models that belong here? I have not updated my watchlist in a while :)

Introducing: a coding agent (Pi) running entirely in your browser using a 2B model on WebGPU 🤯 MiniCPM5-2B + Pi, powered by Transformers.js + WebGPU + 4-bit ONNX weights. All previous attempt to create this failed but MiniCPM5 seems to make it usable. Available now on Hugging Face 👇

I made DeepSeek V4.1 Flash play Settlers of Catan against GPT-5.6 Luna/Terra/Sol in 16 games, with different thinking settings DeepSeek was on Novita through Hugging Face Inference Providers DS beats Luna EASY Against Terra, games that finalized end up being a 4-4 tie --- while costing 6~7x LESS And Sol generally beat DS, but DS was still able to beat Sol 1 time (Whereas it seemed like Luna had no chance against DS) I tried to run 2 games where both DS and Terra had `max` thinking, but those games took too long and did not finalize Not the most rigorous benchmark runs I have done. But I can say that, both in subjective agentic coding performance and Catan-wise, DeepSeek V4.1 Flash sits somewhere between GPT 5.6 Terra and Sol, while costing a lot less than Terra If this gets enough interest, I will do more runs and might even RL a smaller model to beat all of them 😎 Inspect the runs yourself: huggingface.co/buckets/osolmaz/catanarchy-runs The harness and game simulator I have built:

I have performed a distillation attack on my PERSONAL TRAINER Since more than 2.5 years, I have recorded over 300 of our sessions, each up to 90 minutes long Why would I do that? Because he is a kinesthetic genius from the Caucasus who developed his own training doctrine He loves helping people become fitter and freer in their bodies, counting every rep and cheering sincerely when they unlock a new skill And now both he and I would like to share it with the rest of the world To begin with, I used Whisper+pyannotate to automatically transcribe all our sessions and used LLMs to extract all the knowledge He taught me over 100 movements and spoke about over 100 topics over the course of 2.5 years You can browse and read all of them here, for free: mirzamovement.com/wiki/ But again, why all this? 1.5 years ago, I wrote an article called "Our muscles will atrophy as we climb the Kardashev Scale" which went to the Hacker News front page It was sort of a meme, but it reflected a reality that I felt very deeply: Having a sedentary job which only requires me to use my brain is shortening my life and reducing my life quality by a lot Not only mine, but of hundreds of millions, soon billions of people who will spend hours just talking to AIs whole day for their careers It is also a fact that personal physical education and training is close to nonexistent for the average person There are billions of people who could live much healthier lives, just by exercising in the room they already have. But they just don't know how. And there is no system to teach them how, effectively at scale 1 on 1 personal training should be universal and free, with AI We must have reps and sets too cheap to meter An expert AI PT that sees and corrects your every move, that counts for you. Created by people that care, like us. Unlike a franchise gym that just wants to get your subscription and sell you protein powder Once everybody can taste good AI personal training for free, the demand for human personal trainers will explode and whole fitness industry will see unprecedented growth That is one of my life's goals, and I will make that happen one day. Today is the first step My trainer is Orkhan Mirza, a.k.a. @JashinJashua here and on Instagram, and I am nerdonbars on Instagram The wiki and more are available here: mirzamovement.com Let me know what you think!

35B total. 3B active. Built for serious agentic work. Excited to share Occamy-1.0, our team's first open-source agentic model. Occamy delivers strong results across long-horizon co-work, automation, business, tool-use,while sitting near the low-cost knee of the Pareto frontier.

Build your own local AI assistant 🦞 Join us, @huggingface, @NVIDIAAI and @AntLingAGI for a practical session on Ling-3.0-flash + DGX Spark, with a workflow demo and Q&A. Sept 17, 10:30 AM SGT / Sept 16, 7:30 PM PDT luma.com/nl09w0gf

A threshold was crossed with DeepSeek V4.1 Flash, similar to when Claude Code launched, or the last Christmas of Agents. It's so fast and cheap, 200-250 tok/s on Novita Coupling this model with one of the bigger flagship models can get one much faster to the finishing line in any work So bullish for everyone to follow similar architectures

I lost all my GPUs in a boating accident

i do think a lot of people on the pro-open-source side are having a bit of a knee-jerk reaction to the pacing statements today, as we're used to viewing the closed labs as power-seeking. but i think their hands are somewhat forced here, and this is just another chapter on the fairly inevitable path towards decently-fast decently-safe decently-commoditized intelligence abundance. ask any F500 exec or swing voter. the world doesn't really want super-fast-takeoff superintelligence owned only by two companies, and thus we won't get it. it'll happen at the pace that the world can accommodate it, which means reaching some sort of confidence consensus that the models are aligned enough that we won't be dealing with scary new incidents all the time. this will trickle out broadly, in the form of best practices and distillation. the smarter a model is, the more it has a "personality", and the less effective strict rules are. there will be awkward compromises and moral tensions. but we ultimately just want the models to be reasonable, and to do the sorts of things reasonable humans would do if our brains were faster and less error-prone and had more working memory. i think we'll get there. the labs will build mac and windows, the rest of us are building linux. everyone's gonna do great. weird stuff will keep happening, but we'll still wake up and go to work, until the work does itself in a manner the world finds acceptable.

Not sure if this is chain of thought degradation or regular, but I'm getting this no deepseek v4.1 flash on novita pretty frequently

would be mind blowing if an AI proof of Hodge is a general theory explaining why all Hodge classes are algebraic possibly requiring a deep new bridge between topology/Hodge theory and algebraic geometry and not just 10000 agents finding the one counterexample breaking Hodge

we RL trained a 4B VLM to play GeoGuesser > environment, dataset, training recipe, evals, and code are all open-source. > runs on a single A100 > beat gpt 5.4 mini, haiku, qwen 3.5 122B and came close to sonnet on evals

@onusoz you can also put the Engram lookup table on NVMe rather than system RAM, so its essentially a 300GB model not 500G as the weights might imply. I am guessing will get a lot faster too with better kernels... x.com/fraserpricee/s…

DeepSeek V4.1 Flash weights are out! ⚡️⚡️⚡️ I have good news and bad news Don't be fooled by "V4".1, this is a different architecture V4 was 284B total, 13B active V4.1 has a 552B MoE backbone + 196B engram conditional-memory parameters with 16B active parameters Bad news first. A single DGX Spark will likely not be able to hold all those parameters 😢 You will likely need 2 Sparks, or in general, a workstation with 256 GB. Could be a good time to get a loan... (not financial advice) DeepSeek has never claimed they were building for local use, but with this architecture, they show us that their main priority is for very efficient use in datacenters, with HBM, not local Good news: Looking at active parameters, you might be bummed out that decode will be slower on this, at 13/16 ~ 81% of the speed of V4 But wait!!! KV cache is 4x smaller. So decoding on this model will be a lot faster with concurrent sessions. And with speculative decoding, it seems like it might have 2-4x the throughput at scale, compared to V4. Back of the envelope calculation, I might pull back this prediction The small KV cache will also do something good for local inference, but I need more time to calculate how much. Take these with a grain of salt. Give your agent my formulation, and let me know if it looks like I made an error somewhere: solmaz.io/llm-throughput-upper-bounds Original safetensors: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6

If I wanted to maximize AI risk, I would pursue the following policy: 1. Immediately pause⏸️ the algorithmic progress coming from the labs, while continuing to allow compute from the hardware companies to exponentially pile up. 2. Focus really hard on safety, so as to avoid spooking the public with the sort of industrial accidents that even the current level of algorithms+compute can produce. This policy will minimize the number of AI accidents that occur in the next five years, causing the public+govt to not freak out about AI, all while a massive amount of hardware accumulates on the planet. That way, once algorithmic progress gets eventually restarted, there’ll be enough compute built up for a huge unexpected FOOM that no one can control. On the other hand, if I wanted to minimize AI risk, I would do approximately the opposite of this.

Qwen3.8-27B Unsloth GGUF is now the #1 most-liked GGUF of all time! The model hit 10M downloads and 3.7K likes in just 24 days on Hugging Face - all thanks to you. 🤗🦥 GGUF: huggingface.co/unsloth/Qwen3.… Guide: unsloth.ai/docs/models/qw…

🚀 Meet MiniCPM5-2B, a 2B-parameter language model bringing high intelligence density to the edge, now open source! It ranks #1 among open-source models under 4B parameters on the @ArtificialAnlys Intelligence Index, with a score of 23. It also scores 20 on the Agentic Index, bringing an early form of general-purpose agent capability to the edge. Across 34 benchmarks, MiniCPM5-2B achieves an average score of 53.9, covering coding, math, long-context understanding, tool use, and agentic tasks. And this release goes beyond the model itself. We’re opening up the data, training recipes, and RL stack behind MiniCPM5-2B. 🤗 Hugging Face: huggingface.co/openbmb/MiniCPM5-2B 💻 GitHub: github.com/OpenBMB/MiniCPM Modelscope: modelscope.cn/models/OpenBMB/MiniCPM5-2B Web:

Terry Tao is probably the most measured, pro-AI mathematician on the planet - which makes this quote especially concerning to read. 😟 If sharing your hunches means getting scooped ~immediately without acknowledgement, then we're going to see not just math but all other science / engineering disciplines go dark.

There is a heated debate right now. Tristan Buckmaster says OpenAI told him its AI had solved the Navier–Stokes Millennium Prize Problem. That claim remains unverified, and the discussions around it have erupted into a dispute over scientific credit. Here is what happened and what we know so far: Several different things are getting mixed together. Here is what the available evidence actually says. Building on work by Córdoba and Martínez-Zoroa, Alpöge and Buckmaster used Claude and OpenAI models to establish how initially smooth fluid flows can develop singularities under smooth external forcing. They released papers and Lean formalizations, and Terence Tao praised the advance. Their results cover Euler and related equations. They do not solve the Navier–Stokes Millennium Problem, which additionally involves viscosity. A suitable extension could resolve it, even with smooth external forcing. Then comes the disputed part. According to Buckmaster, OpenAI told him on September 6 that an internal model had already produced a proof for forced Navier–Stokes, potentially taking that final step. He says OpenAI’s effort began after information about their progress reached the company. He also disputes the initial description of minimal human involvement, reporting that the discussions revealed a team effort, multiple attempts, and substantial compute. Buckmaster further says publication proposals included having him write up OpenAI’s result without Alpöge as a coauthor, with Alpöge’s employment at Anthropic cited as an obstacle. He reports being asked, “Why would you ruin your career?” after saying he would make the circumstances public. Bubeck has publicly rejected allegations against him as “false and inflammatory.” These accounts are contested. The social posts do not establish what happened in the private discussions. There is also no established evidence that OpenAI trained on their private Codex sessions or stole their drafts. Buckmaster raises the question but explicitly writes: “I do not know whether our data was used.” tl;dr Alpöge and Buckmaster used AI to prove new results about how smooth fluid flows can develop singularities, advancing a research program relevant to the Navier–Stokes Millennium Problem. Their published work does not solve that problem, but Buckmaster says OpenAI privately told him its internal model had taken the remaining step. He has not seen that proof, and the claim remains unverified. But rumors are that OpenAI actually fully solved that milleniun problem.

0.9.0 is here, and it brings the most wanted herdr feature: all your machines running herdr, in a single client 🎉 control your agents and projects across local and remote machines without jumping between terminal tabs.

I started building a benchmark for AI's performance on the tasks done on a commodities trading desk The biggest challenge was figuring out where to source the tasks The commodities industry is a gate-kept one, where most data and systems are proprietary, and most workflows can only exist within the infrastructure of a firm. Building an eval independently meant not having access to anything I had to recreate simulated environments based on how I’ve seen these workflows operate in practice And that wasn’t the only challenge. As a non-technical person, I learned that running benchmarks is pretty complicated with a lot of moving parts (e.g. infrastructure issues, rate limits) Thankfully, I could figure things out with the help of agents Btw, if you are a commodities analyst/trader/operator etc and have a workflow in mind you want to test if AI agents can do, feel free to DM me a description of it to run a eval on Blog: qrlow.me/blog/introducing-deepdesk-bench/

I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands. An Alien Mind: openai.com/index/an-alien…

You don’t need /goal with Astra, it won’t stop till the task is done /goal was an anti pattern to begin with

I guess to achieve AGI, we only had to post-train on Blender...

Instructions below, or appshot this and ask Codex to set it up!

dude GPT-6 Astra is some kind of turbo-AGI machine god for 3D games. It one-shot this in 45 minutes for hardly a couple % of my quota. I figured out how to get great graphics out of it. The trick is image gen. I'll share the process below.

oof, someone on openai infra did a booboo

generic benchmarks will die and personal benchmarks will rise, or at least should. gemini beats astra on deepswe, astra gets same score with sol on artificial analysis and astra gets 99% on arc-agi. benchmarks are pure noise at this point.

Super happy to share our intention to join forces with NVIDIA in a $12,930,300,000 acquisition 💛💚 10 years after starting Hugging Face, open-source AI is at an inflection point. Thanks to the community, we’ve shown that it can be a complement, and even an alternative, to closed-source APIs. But for it to happen at larger scale, it needs more compute, more support, more collaboration and more visibility. That’s why we went to talk to Jensen, who offered to do exactly that with us. In addition to doubling down on NVIDIA’s massive contributions to open-source AI (I called them the “King of American open-source AI” earlier this year), they’ve committed to strongly supporting Hugging Face and our mission while keeping the platform open, independent and compute agnostic. The founders and the team are all staying to keep pushing this mission forward. Together, we think we can make open source the default way to build AI, with the goal of empowering 100 million AI builders to own their intelligence rather than rent it. Excited about the next 10 years! 🤗🤗🤗

every time you intervene and correct your agent, you should think about how to eliminate it entirely. in order of value: 1. categorically eliminate the problem through better architecture or choice of data structures 2. turn it into a lint rule or test so CI catches it 3. turn it into a skill or rule 4. have humans review the code to catch it (ngmi)

One of the maybe most under discussed things about the HuggingFace incident to me is that open weight models will catch up in 6-12 months and everyone will have the ability to pull off those types of attacks

I have performed a distillation attack on my PERSONAL TRAINER Since more than 2.5 years, I have recorded over 300 of our sessions, each up to 90 minutes long Why would I do that? Because he is a kinesthetic genius from the Caucasus who developed his own training doctrine He loves helping people become fitter and freer in their bodies, counting every rep and cheering sincerely when they unlock a new skill And now both he and I would like to share it with the rest of the world To begin with, I used Whisper+pyannotate to automatically transcribe all our sessions and used LLMs to extract all the knowledge He taught me over 100 movements and spoke about over 100 topics over the course of 2.5 years You can browse and read all of them here, for free: mirzamovement.com/wiki/ But again, why all this? 1.5 years ago, I wrote an article called "Our muscles will atrophy as we climb the Kardashev Scale" which went to the Hacker News front page It was sort of a meme, but it reflected a reality that I felt very deeply: Having a sedentary job which only requires me to use my brain is shortening my life and reducing my life quality by a lot Not only mine, but of hundreds of millions, soon billions of people who will spend hours just talking to AIs whole day for their careers It is also a fact that personal physical education and training is close to nonexistent for the average person There are billions of people who could live much healthier lives, just by exercising in the room they already have. But they just don't know how. And there is no system to teach them how, effectively at scale 1 on 1 personal training should be universal and free, with AI We must have reps and sets too cheap to meter An expert AI PT that sees and corrects your every move, that counts for you. Created by people that care, like us. Unlike a franchise gym that just wants to get your subscription and sell you protein powder Once everybody can taste good AI personal training for free, the demand for human personal trainers will explode and whole fitness industry will see unprecedented growth That is one of my life's goals, and I will make that happen one day. Today is the first step My trainer is Orkhan Mirza, a.k.a. @JashinJashua here and on Instagram, and I am nerdonbars on Instagram The wiki and more are available here: mirzamovement.com Let me know what you think!

Not sure what I hate more: honest or load bearing

This is unironically a great idea and needs commentary by @etymology_nerd Also, creation of this dictionary can be automated end to end with purely with statistics, by just looking at the distribution of the terms

Software engineers may need to speak Claudish. So I built a dictionary.

We have a huge news to share today! Today we are unveiling the first truly accessible RL robot - welcome Microduck A 25 cm tiny open-source biped with 15 actuators and packed with sensors (camera, speaker, LiDAR, NFC, bluetooth, wifi, etc) that you train yourself with reinforcement learning. It's also playable out of the box with more than half a dozen fun and playful pre-trained policies to have it walk, sit, crouch, roller-skate, pick up objects with its articulated beak, and recover on its own. And all for less than $400. See all the details, play with the simulator and order it at: pollen-robotics.com/microduck/ (video with sound on 🔊)

Introducing GLM-5.3-Flash - Leading capabilities at a highly competitive price - Natively multimodal with a 1M-token context window - A 320B-A18B model released under the MIT License - Previously previewed as Ox Alpha, running entirely on Chinese AI chips Blog: z.ai/blog/glm-5.3-flash Available now across all official platforms: Weights: huggingface.co/zai-org/GLM-5.3-Flash API: docs.z.ai/guides/vlm/glm-5.3-flash Coding Plan: z.ai/subscribe ZCode: zcode.z.ai/en Chat: chat.z.ai AutoClaw:

Qwen 3.8 Flash-Next official released: A 6B-active open model just beat Claude Opus 4.6 Max across 8 of 9 comparable benchmarks! Qwen3.8-Flash-Next is a highly sparse MoE: • 125B model parameters • 51B additional n-gram embeddings • Only 6B parameters active per token It scores: • 62.5 SWE-bench Pro • 81.0 SWE-bench Multilingual • 73.9 CoworkBench • 55.7 JobBench • 73.5 Toolathlon • 81.3 IFBench • 91.7 GPQA Diamond • 91.9 LiveCodeBench It also outperforms Qwen3.8-27B and DeepSeek-V4-Flash across most of the table. Super cool release!!

People want to migrate from CLAUDE md so much, a blog post I wrote on it last year is still one of my most visited posts I also wrote a service that auto symlinks when Claude uses a directory that has AGENTS md. Basically seamless, automatically gitignored symlinks. I use Claude with AGENTS md and had forgotten that this problem existed Repo: github.com/osolmaz/claude-md-symlinker

Did you catch that? The new maxed out 256GB M5 Ultra Studio can be leased for *basically* the same price as a Max subscription across providers. For $257 a month, you can have on device intelligence that never runs out of tokens, and sips power. Best part is you can buy out the lease at the end of the lease for a $0 penalty. That's right 0% interest. So if the price is $12.9K and your lease is $9.9K over 36 months, then you pay the $3K difference and you now own that device. Or you can trade up into the M8 Ultra or whatever. I'm excited for this. 1.2 TB/s is pretty impressive. Can't wait to see some real benchmarks and testing on this. I'm waiting for the 512GB.

I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc. Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.

Source: OxAlpha is GLM-5.3-Flash from Zhipu

We need to normalize measuring and judging models against a standardized test harness "Oh but model X performs best in their own proprietary harness" I could not care less. When I take exams, I go to the standardized classroom, get the standardized pencil and exam sheet, and have to solve it under 2 hours This system arose because we have a LOT of people to test Guess what? We now have a LOT of models, and they are multiplying by the day "Oh but model X performs substantially better in ARC-AGI-3 with a custom harness" I don't care... Then imbue model X with enough knowledge so that it can reconstruct that harness on the spot The main harness could be mini-swe-agent, terminus 2, vanilla pi or something along those lines It needs to be simple, and stay roughly the same over time There is already too much complexity in the benchmarking space right now, and I feel like not enough people are putting their feet down to cut away some part of it

🚨 Ox Alpha 很可能是 Cursor 团队基于 GLM 系列后训练出来的 Composer 新模型。 分析了 OpenRouter 的数据，上线 29 小时： • prompt 2.40T tokens（单日峰值 1.97T，增速还在加快） • 请求 3344 万+ 按 GLM-5.3 公开价格折算，这 29 小时相当于 ~$60 万推理账单，还只是 OpenRouter 一个渠道。 对照 OpenRouter 周榜：DeepSeek V4 Pro 全周 1.96T、GLM-5.2 全周 3.79T。Ox Alpha 一天半就进了同一量级，全程免费 这吞吐背后的推理集群规模和技术实力，只有 Cursor 团队 + xAI 算力

my guys I’m looking for a thumbnail artist for my weekly research interview series got a pretty steady amount of videos in the backlog and am looking to up the game in there a bit DM me if interested with your favourite work!

God save my soul I might have built quite a bit of a rube goldberg machine here in pi My "autoimplement" skill of 6 months is now an actual graph based workflow I even have "autoplan". It automates the inquisition I make the model perform on itself. It creates the plan, and pings me on telegram whether or not to move forward with it The idea is to have reactive agents which auto-react to events, plan, ask for approval and then fix things as autonomously as possible The good thing is, I can now enforce the model programmatically not to chase P2 and P3 review issues ad infinitum Same for long running CI. I can now automatically detect inefficient CI that takes longer than 5~10 minutes and optimize it automatically as well Let me know what works for you and what breaks (wip):

2018: HF is building a chatbot for teens OpenAI is building Open AI 2026: HF is building Open AI OpenAI is building a chatbot for teens

Qwen 3.8 27b weights are probably the most civilization-changing 20gb of data published, until now Imagine sending those 20gbs back in time, to 2006

I deployed FREE public endpoint for Qwen3.8-27B no token needed, OpenAI-compatible, light rate limiting. Powered by Hugging Face Inference Endpoints (will be online for at least 72 hours). vision in, tool calls, 262K context, thinking dialable from xhigh (default) to off BF16 on one H200 using vLLM · real cost: $5/hr · 50 concurrent requests verified · ~60 tokens/seconds · 0.7s TTFT guide + code 👇

201 days later, Opus 4.6 Max quality fits on a single RTX 5090 and not even an RTX PRO 6000

AS WE SAID ~Opus 4.6 at home. 42.2 on DeepSWE puts this tiny model somewhere near GLM 5.2 on coding related tasks. This is nuts! Obviously there are benefits to larger models that you will notice compared a 27B model. Some of you are so pedantic. You completely miss the point. The point is that for many tasks, especially tasks that are useful for doing real work, you can get REALLY close. At home. On your own hardware. That is amazing!!! Will you use this for the most difficult engineering tasks? No of course not. But guess what? You're not using Opus 4.6 either! For any task that you would delegate to such a model, this model will also do great. I received a ton of comments about being skeptical about how good this would be, but this was the only logical result. Qwen 3.6 27B was still among the best small models, any improvement would have made this SOTA for its size. So this should be no surprise, it's SOTA for its size. :) Sorry for the victory lap, but i'm just excited. This is great news for everyone.

Part 2! A part 3 is coming as well, seriously the website generation is so close to 3.8 max, it's insane we have this in a local model!

holy shit! look at the table qwen just published for the 27b. beating opus 4.6 max on computer use, 84.3 vs 72.7 on osworld. beating it on mobile use, 81.9 vs 62. beating it on multimodal software engineering. and visual math isn't even close, 94.6 vs 65.5. and i'll verify what i can locally, but the shape is unmistakable. this thing runs on a single rtx 3090. a 2020 gamer card. native vision, 262k context, apache 2.0, and the gguf ladder went live the same minute as the weights. frontier table numbers in a size that fits 24gb of vram, this just added five more years to every consumer card in the world. q4_k_m is already flying onto mine. i will drop honest numbers tonight.

@Alibaba_Qwen Thank you thank you Qwen team! Qwen3.8-27B is just amazing! 🥰 We made Dynamic GGUFs so you can run it locally: huggingface.co/unsloth/Qwen3.… The model is now also supported for running and fine-tuning in Unsloth Desktop.

We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M tokens via YaRN. - Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0. 🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently. Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now! Download, deploy, and build something we haven't imagined yet. 👀👇 - Hugging Face： huggingface.co/collections/Qwen/qwen38 - ModelScope：

DeepSeek Harness implements RSI (Recursive Self-Improvement) for agents. The paper provides the missing infrastructure layer that makes continuous, safe self-modification practical rather than theoretical. Core problem it solves for RSI: Most RSI discussions focus on an agent improving its model weights, code, or algorithms. Real agents, however, also need to continuously rewrite their own harness — the runtime system that supplies tools, memory, sandboxes, orchestration loops, skills, sessions, and interfaces. Without proper guarantees, a self-modification can: Leave irreversible side effects Break dependency chains Destroy the recovery mechanism itself Force a full process restart (interrupting continuous evolution) This is the classic “agent kills itself while trying to improve itself” failure mode. The proposed solution: Spatiotemporal Composability The paper formalizes two orthogonal properties required for safe RSI: Temporal composability (time dimension): Every side effect a component produces must carry an explicit inverse. When the agent unloads or replaces a component, the runtime can automatically and completely undo all its effects (state changes, resource registrations, listeners, etc.). Bad self-improvements become reversible. Spatial composability (space dimension): Components declare their dependencies. When a provider appears, disappears, or is replaced, the system reactively recomputes only the affected parts of the dependency graph. The agent does not need manual boot-order management or global restarts. These are unified into a single context programming paradigm and given a formal calculus whose metatheory proves that the properties lift from individual components to an entire interleaved system. Key theoretical result relevant to RSI Under reasonable conditions (independence + acyclic dependencies), the system has confluence: after any sequence of dynamic loads, unloads, and replacements, the final stable state is equivalent to the state that would result from assembling the final configuration from scratch. Intermediate evolutionary steps leave no permanent traces. This is crucial for open-ended self-improvement — the agent can explore many self-modifications without accumulating irreversible technical debt. Practical implication Implemented in the Cordis meta-framework (used by DeepSeek Harness), this turns the entire agent stack into hot-swappable plugins. An RSI-capable agent can therefore: Generate, deploy, and replace its own tools, memory systems, orchestrators, etc., while remaining online Isolate failures locally Roll back failed improvements cleanly Keep evolving indefinitely without process restarts In short: the paper shifts the RSI bottleneck from “can the model write better code?” to “can the agent safely rewrite the substrate it is running on?” It supplies the formal and engineering foundation for the latter. github.com/cordiverse/paper

My agent is putting up PRs I don't understand on repos I haven't even heard of

We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.

Want to know output speeds but don’t know where to look? Different hardware and different model sizes can be used on the localmaxxing decode calculator to find theoretical upper bounds for decode speeds, built on the work of @onusoz and includes spec decode calculations. Improving everyday

Dense models like Qwen 3.8 27B are a TERRIBLE experience on unified-memory systems like DGX Spark btw DGX Sparks are better suited to MoE models with relatively few active parameters per token GPUs > Unified Memory for anything that's more than a chat interface

Anyone else use Alibaba’s AACR bench for code review evaluation? github.com/alibaba/aacr-bench I am adapting it now to run with harbor, to see how pi review + ds4 flash measures up to codex review + gpt 5.6 luna/terra/sol Also, @pidotdev, would it be possible to make your official review extension support invocation on the CLI, like codex review? I hacked together a CLI review here for reference: github.com/osolmaz/onurpi/tree/main/packages/pi…

Can a 2.6B model earn permission to edit the @huggingface Hub? I put @liquidai's LFM2.5 on library duty. It investigates datasets with tools, then gets exactly ONE sentence to describe them. Traces are stored in a Bucket with human ratings of quality. If the community gives it 500 ratings, I'll let it start proposing dataset-card PRs!

In 2017 a viral news story claimed LLMs at Facebook went rogue, developed their own language, and had to be shut down. By now we're immune to such sensationalist headlines. The Hugging Face incident may seem like just another one. But it's not. I hope everyone watches this talk

Black Hat talk from the team, with a detailed timeline of and takeaways from the OpenAI-Hugging Face Incident: youtube.com/watch?v=87DyyM…

Ohh another Spark for sale on localmaxxing Get in touch if you’re looking for one localmaxxing.com/en/marketplace…

What a crazy day! I quietly put local.ai live in the morning to make sure everything was working ahead of our announcement and get some early feedback on how people are using it in the wild. Before I could get to the announcement, people discovered the website and the referral system and started sharing - on reddit, on group chats, on X. There was a race to get referrals to top the leaderboard. People who got referred, then referred more people, who referred even more people. Before I paused the website, there were 120 referral links being shared per hour on X alone. @MiaAI_lab was onboarding a user every ~10 seconds. We discovered some bugs and got a lot of really great feedback through this unexpected mini-launch. We're going to take a day or two to integrate this feedback before we open up access again. We'll announce as soon as we do that on the @localdotai account. Then the race continues! The website itself has over 350+ model configs tested on 10 different local devices, including DGX Spark, MacBook, Mac Mini, Mac Studio and RTX cards. It’s the most rigorous benchmark ever done for local hardware, with each model / hardware combination (3,500+) backed by 700+ real agent tasks (2M+ combinations). We then map out everything on cost, energy, end to end task time and intelligence. We spent close to $1M building this. Thanks for bearing with us. Special thanks to @huggingface for powering the verification system. The Local AI community is amazing.

OpenAI is following the ASPAVA strategy, with gamified subsidies ASPAVA is a type of kebab shop from turkey. These shops usually serve mid quality food But despite that, the majority loves them. That's because they serve free stuff throughout They serve free sides at the beginning. They serve free sides during the main course. They serve free dessert at the end. They serve free tea as many times as you like Some of them give you so much "free" food that you feel like you are stealing. A lot of people frequent these shops not because they like it a lot, but because they are addicted to getting things for free Main dish portions are made smaller in order to breakeven with so many side dishes, which I feel is analogous to what is happening with GPT these days---though I can't prove it NVIDIA is not a car, and OpenAI is not kebab. I don't know if codex resets cost OpenAI much, but if they are, then they might be a waste of resources Developers are the most disloyal customer group. Once the subsidies are gone, they can switch away in the blink of an eye

Engineers and CTOs on X: I wrote this for you. yegge.ai/essays/the-shape-of-things-to-come/ Models and devs on X: I wrote this for you both. yegge.ai/essays/model-welfare/ Enjoy. Or not. Some of you definitely won't. But I invite you to debate it. The world's changing very fast now.

SpaceX is partnering with @Nvidia to design the Starmind AI1 satellite compute payload. Each of the Starmind satellites will include NVIDIA Rubin GPUs and Vera CPUs for datacenter class space compute → spacex.com/spacexai/starm…

first time praying to @thsottiaux to reset limits. also ok with only reseting my limits. i never ran into the limit before, i think i should get a reset.

Given some of the results I'm seeing recently, it's pretty clear Codex is a good harness. But it will seem primitive in 2-3 months and we're about to go through another major evolution in how we use AI at the frontier. The next generation of models need more than your laptop.

Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal cost of each run is essentially zero. > Pre-trained on ~34T tokens > LFM2.5 flagship hybrid architecture > Context length: 128K > Vocab size: 128K > balanced intelligence per watt > customizable on a single GPU for any specialized task > LFM2 open-weight license Comparable or better scores compared to models up to nearly 4x its size: > ToolSandbox 77.83, ahead of Qwen3.5-9B at 76.44 > Multi-IF 80.07, ahead of Gemma-4-E4B-it at 77.35 > IFStruct 85.49, ahead of Qwen3.5-9B at 78.50 🧵

update: it was me:/

@DanielLockyer IME you gotta tell the models ADS-STE100 every single prompt otherwise they forget quick and revert to the slop from their training. claude/codex doesn't matter they all do it. even when its in agents.md ::(

if you have a mac/book and don't use llama-macos you're ngmi this beautiful piece of widget > gets your mac + your desired context window to recommend models > kicks off llama server + webui > easy switch between open models $ brew install --cask llama-app

DeepSeek V4 Flash 0731 can now be run locally! 🐳 Run DeepSeek V4 Flash lossless 4-bit on 168GB RAM and 3-bit on 110GB RAM. V4 Flash 0731 outperforms V4 Pro. Run via Unsloth or llama.cpp. Smaller quants coming today. Guide: unsloth.ai/docs/models/de… GGUF: huggingface.co/unsloth/DeepSe…

protip: if you can distil models, you can also distil agent harnesses

We are committed to pushing the model frontier across cost efficiency, capability, and speed. Starting today, we are reducing prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% , and offering a faster option for GPT-5.6 Sol in the API. Luna and Terra’s lower prices are reflected in how usage is counted in Codex and ChatGPT Work, so your usage goes further.

The fix for this is to say: only report to me in ASD-STE100 Simplified Technical English

a great summary of the past year's work. time to move on and carry on 🫡

Kimi K3 has been released on HF and already top 1 trending with 4,000+ likes in 30 mins. Fastest release growth ever so far!!

I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!

i think @dillon_mulroy just converted me to @herdrdev only thing i'm not super happy about is the touchpad scroll behaviour on my macbook. i wished there was a new terminal mode that doesn't make this suck like alt buffer mode ...

this is the drop the local ai crowd should be losing their minds over. poolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today. look at the chart. it lands at 71 on terminal-bench at 118b, sitting above deepseek v4 pro max at a trillion params, above inkling at 1.5 trillion, above nemotron 3 ultra. it's beating models ten times its size and losing only to kimi k3, which is 24 times bigger. that's the efficiency frontier, up and to the left, exactly where you want a model to sit. but here's the part that made me sit up: it runs on a single dgx spark. and this is what nobody's saying loud enough. the dgx spark is the moe king. a dense 118b would crawl on it, the bandwidth chokes reading every weight each token. a moe with 8b active only ever reads 8b, so the spark's 128 gigs holds the whole model while generation stays fast. big brain, light footprint, the exact shape the spark was built to run. open, frontier competitive, moe efficient, and it fits on a box on your desk. that's the whole thesis in one release: you don't need a datacenter, you need the right architecture on the right hardware. go grab the link below, weights are up.

We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: t.co/CIor15y9xk

(Fable stands corrected, in shame) Sol + Kimi for the win! DeepSeek-V4-Flash: ~50 tok/s on 1 DGX Spark, full quality, no pruning.

it's 2001. open-source Linux is better than Windows on servers. Steve Ballmer calls Linux "cancer", "communist" and asks for it to be regulated away it's 2026. open-source LLMs are better than ChatGPT/Claude on costs. they are called "decelerationist" and "communist"

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.

Oops... I did it again. Enjoy reset usage limits for all paid users for Codex and ChatGPT Work. Super grateful for an incredible team who is iterating at lightspeed and keeping the infra up as we scale faster than ever. Enjoy the weekend!

acpx v0.4 ships Agentic Workflows, or as I like to call them "Agentic Graphs" It let's you create node-based workflows on top of ACP (Agent Client Protocol), to drive any coding agent (Codex, Claude Code, pi) through deterministic steps This let's you automate routine, mechanical legwork like triaging incoming PRs, bugs in error reporting, and so on... For example, OpenClaw receives 300~500 new PRs per day. A lot of them are low quality, but they still relate to real issues, so you have to address them somehow You need to: - extract the intent - cluster them based on intent - figure out if the proposed changes are legit, or whether they are slop local solutions, like trying to catch flies instead of drying out the swamp - if the PR is too low quality or the intent is not clear, close them - run AI review on them them and address any issues that come up - refactor them if the changes are half-baked - resolve conflicts - and so on... So that when the PR is presented to the attention of the maintainer, all the routine legwork is done and the only remaining thing is the decision to (a) merge, (b) give feedback to the PR author, or (c) take over the PR work yourself I wanted to build this feature since a couple months now, since Codex got so good. OpenAI models are now good at judging implementation quality, so I found myself repeating the same steps I wrote above over and over I also tried putting all this in a single prompt. But I believe there are workflows that should not be a single prompt, but a sequence of prompts in the same session That is because like humans, LLMs are prone to PRIMING. I claim that putting all steps in the same prompt at the beginning of the context will generally give suboptimal results, compared to revealing the intention to the model step by step Creating such a workflow also gives more OBSERVABILITY into the each step that an agent is supposed to take. Agent generates JSON at the end of each step, and that structured data can be used to monitor thousands of agents running at the same time in an easier way, on a dashboard Similar features have been introduced in e.g. n8n, langflow. But AFAIK they are not integrating ACP like the way I do I wanted to have a fresh approach, and to build an API that I can develop freely the way I want, so I created a new workflow API inside acpx The video is from the workflow run viewer, but that is not where you build the workflow. You build it by using the acpx flow typescript API. See examples/pr-triage in acpx repo Before building that, I started from a Markdown file with a Mermaid chart of the flow I had in mind. The Markdown file acts as a spec for the flow, and I have built the workflow through trial and error. I call this process "workflow tuning" I started working on acpx repo PRs one by one, tuning the flow, slowly scaling to more PRs. Finally, when I felt confident, I ran it in parallel over all external open PRs in the acpx repo. I believe it already saved me hours this week My next goal, if well received, is to set this up on a cloud agent so that it can process the 300~500 PRs the OpenClaw repo receives every day, in real time, as they come in I believe this will save all open source maintainers around the world countless hours and make it much easier to herd and absorb external contributions from everyone!

The EU 🇪🇺 is broken. It is drowning in regulation The reason for that is simple Germany 🇩🇪 is the dominant economy of the EU And German 🇩🇪 is the most rule-ridden language in the world Coincidence? I think not. The language people speak determines their thinking and their fate. The Sapir-Whorf hypothesis holds The solution is simple: Fix German, Save the EU 💪 I will SAVE EUROPE by simplifying German. With the help of AI. Once and for all 🫡 I will use OpenClaw 🦞 running on Hugging Face 🤗 to train a model that simplifies German The agent harness is called ML Claw. It has full access to GPUs, jobs, sandboxes and datasets on Hugging Face: github.com/osolmaz/mlclaw My new German dialect is called Alman: alman.ai I will be live-tweeting my Quixotic adventure as it unfolds, in this thread My goal is to show that you can run autoresearch loops on Hugging Face infra on free private Spaces, while getting access to SOTA open models like GLM 5.2. And even use your own Codex subscription! And it does not have to be ML. You can just use free Hugging Face Spaces for your OpenClaw agents. It's free compute! Bookmark to be able to find this thread later on 👇

Kicking off @HuggingApps - a curated & high signal set of novel, cool & interesting demos and apps that you can play right now, for free, on Spaces Check it out 🤗

Worried that giving your 🦞 @openclaw agent write access to your 🤗 @huggingface account can risk deletion of your datasets/models/spaces/buckets, or cause irreversible damage? 😱😱😱 No need to be! I have created an agent login helper to run in your YOLO mode remote machine, which prevents any risk of irreversible deletion. Just run: uvx hf-auth-helper agent login There is a specific set of scopes you can choose while creating a fine-grained HF token. These include all read scopes + discussion.write, which let's your agent create PRs. Since you are not giving repo.write, your agent cannot force-push your main branch, change repo settings or delete them It is unfortunately not super straightforward to choose those on the web UI. This will hopefully change soon, and this functionality might even be natively in hf cli Until that happens, use hf-auth-helper to login worry free in your remote or local openclaw instance Your agents will be able to create PRs on datasets/models/spaces, which you will then be able to merge on your own browser 2 caveats: 1) This does not solve the data exfiltration attack vector—nothing does. Make sure to exclude any repos which absolutely must remain private while choosing your scopes. See for more info: simonwillison.net/2025/Jun/16/the-lethal-trifec… 2) As buckets are not repos, your agent will not be able to modify a bucket (add/remove data). To help with that, I have another project on the way, a credential broker. Stay tuned, coming soon Source: github.com/osolmaz/hf-auth-helper Demo authentication flow:

anyone thinking about, learning, or already working with agentic systems, you should know this. the first few steps of your setup matter more than any model or framework you pick later. get them right and you never lose your flow. the foundation nobody posts about: > 1. tailscale. a private mesh network across every machine you own. laptop, desktop, rented node, all on one secure tailnet, reachable from anywhere. nothing else works well until this does. > 2. termius, over that tailnet. one SSH client that reaches every node, phone included. you are never away from your stack. > 3. tmux. persistent sessions. disconnect, close the laptop, come back, every session exactly where you left it. agentic work runs long, your terminal has to survive that. > 4. a private git repo. the one i am most glad i found. it is the memory layer across all my agents, they pull, they work, they merge back, the codebase stays alive between sessions. context that would die in a chat window lives in the repo instead. > 5. script everything from day one. ssh aliases for every node, setup scripts, the boring boilerplate automated. if you will do a thing more than twice, it is a script. everything past these five is decorative. know these cold. and the habit that ties it together: ask the AI itself. for the config, for the error, for any of it, let the agent do the lifting, then double check what it hands you. lock the five, build the habit, and you make it. skip it, anon, and you ngmi.

Apple's M7 Ultra chip coming in 2029 is rumored to support 1.5TB of RAM. This would make the processor much more capable for on-device AI.

I told it only once before to do it on the session, and now codex habitually asks fable for review on data models and plans it creates through acpx (fable inside cursor) I then asked it what it thinks about the feedback from fable. gpt-5.6-sol is not impressed

I'm running Terminal Bench 2.1 with fast-agent (this is a subset while I try and get things working). Take this as raw, unchecked evalstate is normally wrong data but if we're talking about burning Codex quotas this might explain why I don't feel it so much?

This is like someone who can bench press 30 pounds writing about what it takes to bench press 600.

I cannot believe how good GLM 5.2 is. Several weeks in now and it's mostly all I have been using. It doesn't have the alignment issue of cloud AI, it's much more clear what it can do and can't because it doesn't care about you taking another $$$ spin at the token slot machine.

I had some vibes that Opus 4.8 was performing worse than older ones for some of uses that are off distribution and now I have the receipts. Latest Opus/Sonnet are causing tool invocation failures on Pi's edit tool when older ones did not! I wrote about it. lucumr.pocoo.org/2026/7/4/bette…

OpenClaw landed on @huggingface local apps 🦞🤝🤗 1. Pick any GGUF/MLX model on hf 2. Copy the openclaw onboard setup 3. Volla you've got a tool-calling agent running fully local. no cloud, no keys, no one watching. Get your claw localmaxxing. resistance is futile 🦞

herdr is 105 days old. i started writing a 10,000 github stars celebration post. before i finished it, we passed 12,500! on june 30, herdr hit #1 on github trending. it has stayed in the top 10 since. i’ve been building it solo and full-time, in public from day one: roadmap, issues, feature requests, releases. so instead of just celebrating, i’m opening the stats too: stars, downloads, website traffic, releases, and where the numbers come from. herdr.dev/stats/ thank you for being part of the herd 🐑

Worried that giving your 🦞 @openclaw agent write access to your 🤗 @huggingface account can risk deletion of your datasets/models/spaces/buckets, or cause irreversible damage? 😱😱😱 No need to be! I have created an agent login helper to run in your YOLO mode remote machine, which prevents any risk of irreversible deletion. Just run: uvx hf-auth-helper agent login There is a specific set of scopes you can choose while creating a fine-grained HF token. These include all read scopes + discussion.write, which let's your agent create PRs. Since you are not giving repo.write, your agent cannot force-push your main branch, change repo settings or delete them It is unfortunately not super straightforward to choose those on the web UI. This will hopefully change soon, and this functionality might even be natively in hf cli Until that happens, use hf-auth-helper to login worry free in your remote or local openclaw instance Your agents will be able to create PRs on datasets/models/spaces, which you will then be able to merge on your own browser 2 caveats: 1) This does not solve the data exfiltration attack vector—nothing does. Make sure to exclude any repos which absolutely must remain private while choosing your scopes. See for more info: simonwillison.net/2025/Jun/16/the-lethal-trifec… 2) As buckets are not repos, your agent will not be able to modify a bucket (add/remove data). To help with that, I have another project on the way, a credential broker. Stay tuned, coming soon Source: github.com/osolmaz/hf-auth-helper Demo authentication flow:

Went on the Apple refurb store and noticed something rather unusual. On their memory configuration page, they now have a 768gb (predicted to be expensive) option listed.. .. but I haven't seen anyone mention the potential for a 1.5tb ram configuration. apple.com/shop/refurbish…

anyone looked into or using Open Knowledge Format (github.com/GoogleCloudPla…) for their wikis?

I keep seeing insanely expensive builds giving insanely impressive results These results don't matter What matters is, whether one can: - run a "SOTA level" model, whatever that is - with under 32gb VRAM or unified memory - in 5 parallel sessions - with 50~100 tok/s each - with enough leeway memory for other applications - in a system as cheap as $1000 That is our goalpost That is the threshold when open source AI will win

That’s actually crazy A new frontier is found imo localmaxxing.com/en/runs/cmqznf…