Image created with Flux Pro v1.1 Ultra. Image prompt: Ornate showgirl glamour in orange-and-teal tones, dazzling curtain backdrop opening to reveal radiant gears and code, stylized text “OpenSource” in glowing marquee script across the opening curtain; spotlit, dramatic contrast, vintage grain, cinematic, high-detail
OpenAI / America is still ahead in the race”” -> no There is no western open-source model that beats or ties the best chinese open-source models.”” / X https://x.com/scaling01/status/1952900225120780705
Did yesterday’s release shift the needle in the open vs. closed debate? Today in @ReedAlbergotti’s newsletter https://x.com/fdaudens/status/1953147586312872057
I signed this because, despite worrying about misuse of open models more than most, I would like that to be the bottleneck rather than “”is it beneficial to big companies commercially/reputationally etc.”” There are many benefits to the US investing here. https://x.com/Miles_Brundage/status/1952400404668657966
RT @natolambert: America needs to take open models more seriously. This summer the early lead in open model adoption of the US via Llama ha…”” / X https://x.com/ethanCaballero/status/1952459460703834392
The relative failure of Llama 4 turned out to be very consequential to the AI landscape. It led to the shifting the locus of open weights development to China, a move towards closed models as companies running local Llama couldn’t continue to upgrade, & big talent wars in the US.”” / X https://x.com/emollick/status/1951433537485500476
The US now likely has the leading open weights models (or close to it)… … but the real question is whether this is a one-off situation from OpenAI, in which case the lead will evaporate quickly as others catch up. But also unclear what their incentives are to keep updating.”” / X https://x.com/emollick/status/1952836130958917894
Why open-source AI became an American national priority | VentureBeat
China’s ByteDance just released an LLM-based agent for general purpose software engineering tasks. Trae Agent comes with an interactive CLI that can execute complex workflows using simple English prompts. It works with OpenAI and Anthropic API. 100% opensource. https://x.com/Saboo_Shubham_/status/1942047679758151783
America needs to take open models more seriously. This summer the early lead in open model adoption of the US via Llama has been overtaken by Chinese models. With The American Truly Open Models (ATOM) Project we’re looking to build support and express the urgency of this issue. https://x.com/natolambert/status/1952370970762871102
very excited by the ATOM project”” / X https://x.com/finbarrtimbers/status/1952401883391520794
Every tech company can and should train their own deepseek R1, Llama or GPT5, just like every tech company writes their own code (and AI is no more than software 2.0). This is why we’re releasing the Ultra-Scale Playbook. 200 pages to master: – 5D parallelism (DP, TP, PP, EP, https://x.com/ClementDelangue/status/1952048356710039700
Open models by OpenAI | OpenAI https://openai.com/open-models/
holy shit get ready for a hallucination fiesta with gpt-oss https://x.com/scaling01/status/1952781018554933261
It seems the closed-source vs open-weights landscape has been leveled. GPT-5 is just 10% better at coding than an open-weight model you can run on a consumer desktop and soon laptop. If Anthropic cannot come up with a good model, then we will probably not see AGI for a while.”” / X https://x.com/Tim_Dettmers/status/1953521836299350494
8.6% of the world’s population uses ChatGPT weekly…”” / X https://x.com/emollick/status/1952389693502370198
A hypothesis: gpt-oss is trained entirely on synthetic data, from pre-training to post-training. The approach enhances safety and helps smaller models achieve better performance.”” / X https://x.com/huybery/status/1952905224890532316
attention is 0.84% of gpt oss, intelligence is stored in those 99.16% mlp layer, attn is key to unlock it https://x.com/shxf0072/status/1953143243992166849
BREAKING: OpenAI just released two open-weight models: gpt-oss-120b and gpt-oss-20b. The 120B model is on par with o4-mini on reasoning benchmarks and can run on a single 80GB GPU. The 20B model achieves similar results to o3-mini and can run on edge devices with 16GB of https://x.com/rowancheung/status/1952777754904072566
curious about the training data of OpenAI’s new gpt-oss models? i was too. so i generated 10M examples from gpt-oss-20b, ran some analysis, and the results were… pretty bizarre time for a deep dive 🧵 https://x.com/jxmnop/status/1953899426075816164
Everyone is sleeping on AMD for local models – gpt-oss 20B running on an AMD GPU @ 52 tok/sec in a <$1000 laptop https://x.com/dzhng/status/1953132623280165193
gpt-oss for entirely local tool use:”” / X https://x.com/gdb/status/1952802157956350221
gpt-oss https://gpt-oss.com/
gpt-oss is a big deal; it is a state-of-the-art open-weights reasoning model, with strong real-world performance comparable to o4-mini, that you can run locally on your own computer (or phone with the smaller size). We believe this is the best and most usable open model in the”” / X https://x.com/sama/status/1952778518225723434
gpt-oss is out! we made an open model that performs at the level of o4-mini and runs on a high-end laptop (WTF!!) (and a smaller one that runs on a phone). super proud of the team; big triumph of technology.”” / X https://x.com/sama/status/1952777539052814448
gpt-oss-120b & gpt-oss-20b Model Card | OpenAI https://openai.com/index/gpt-oss-model-card/
GPT-OSS-120B casually calculating the product of two random 30-digit numbers. without any tools, just 18k tokens https://x.com/scaling01/status/1952892387539259455
Just released gpt-oss: state-of-the-art open-weight language models that deliver strong real-world performance. Runs locally on a laptop! https://x.com/gdb/status/1952780717638942910
RT @ggerganov: Llama.cpp supports the new gpt-oss model in native MXFP4 format The ggml inference engine (powering llama.cpp) can run the…”” / X https://x.com/ggerganov/status/1952978670328660152
RT @OpenAIDevs: Student credits for gpt-oss With @huggingface, we’re offering 500 students $50 in inference credits to explore gpt-oss.…”” / X https://x.com/reach_vb/status/1953010091377958984
Frontier models, capable of agentic reasoning, can now run on your Macbook Pro 🧑💻 @OpenAI’s release of GPT-OSS 20B and 120B are the biggest releases in open-source this year. Build agentic workflows with @llama_index that run 100% locally! Huge props to @LoganMarkewich and https://x.com/jerryjliu0/status/1952883595787239563
GPT-OSS models seem to be slopmaxxed on math/coding and reasoning – they are great at that but they completely lack taste and common sense at least that’s my vibe so far”” / X https://x.com/scaling01/status/1952881329772564764
I think gpt-oss was always expected to be put in an agent harness that uses search for all its world knowledge. Ive always argued this is not a valid replacement, the rich connections it builds from actual backprop on the worlds knowledge – not just facts, but the aggregate”” / X https://x.com/Teknium1/status/1953230352568467761
I was just about to make a post that GPT-OSS-120B is nontheless an overall good for the very low end. But I honestly don’t know what it is good at, except benchmarks. Coding seems to suck, creative writing is terrible… So it’s just a math model? https://x.com/scaling01/status/1953047913954791696
I’m thrilled @OpenAI has released two open weight models. Thank you to all my friends at OpenAI for this gift! I’m also encouraged that from my quick tests gpt-oss-120b looks strong (though we should still wait for rigorous 3rd party evals).”” / X https://x.com/AndrewYNg/status/1952838045235126510
i’ve spent the last couple hours talking to gpt-oss and can safely say it’s unlike any model i’ve tested one second it’s coding for me at a professional level, the next it’s making up basic facts and clinging to them no matter what i say something very strange is going on”” / X https://x.com/jxmnop/status/1953216881361600729
I’ve written the full story of Attention Sinks — a technical deep-dive into how the mechanism was developed and how our research ended up being used in OpenAI’s new OSS models. For those interested in the details: https://x.com/Guangxuan_Xiao/status/1953656755109376040
ICYMI: you can vibe test the latest gpt-oss models on gpt-oss[.]com 💥 We partnered with @OpenAI to bring easy access to the model right down to a browser near you! https://x.com/reach_vb/status/1953041435999010916
Introducing gpt-oss | OpenAI https://openai.com/index/introducing-gpt-oss/
Is it over for gpt-oss ? What are these Aider Polyglot scores? https://x.com/scaling01/status/1952780629772321257
It’s looking bad bois.. Aider Polyglot results for GPT-OSS-120B: 41.8% for comparison: Kimi-K2: 59.1% DeepSeek-R1: 56.9% Qwen3 32B: 40.0% https://x.com/scaling01/status/1953047534122713130
Our new @OpenAI open models https://x.com/polynoamial/status/1952778238368887184
Thank you @OpenAI for open-sourcing these great models! 🙌 We’re proud to be the official launch partner for gpt-oss (20B & 120B) – now supported in vLLM 🎉 ⚡ MXFP4 quant = fast & efficient 🌀 Hybrid attention (sliding + full) 🤖 Strong agentic abilities 🚀 Easy deployment 👉🏻”” / X https://x.com/vllm_project/status/1952784530466849091
The gpt-oss models have been post-trained to use two specific first-party tools: 1. a web browser that can search, read pages, follow links, and cite sources 2. an interactive python notebook This will give gpt-oss based agents super powerful capabilities out of the box! https://x.com/corbtt/status/1952810876165312805
We fixed some issues for @OpenAI’s gpt-oss model! 1. Jinja template has extra \n s, didn’t parse thinking sections + tool calling wasn’t rendered correctly 2. Some versions miss <|channel|>final -> this is a must! 3. F16 infs: use F32+BF16! We made a few free Colab notebooks as https://x.com/danielhanchen/status/1953901104150065544
Well, it took just 2 hours for OSS-GPT to hit #1 on @huggingface. Don’t remember seeing anything rise that fast! https://x.com/fdaudens/status/1952814865795698954
🚨 It’s official: OpenAI’s gpt-oss-120b & gpt-oss-20b just landed on Hugging Face! Brand new open-weight LLMs ready for anyone to try, fine-tune, and run anywhere. Here’s what makes this drop a big deal: https://x.com/fdaudens/status/1952781183575593234
And just like that, @OpenAI gpt-oss is now the number one trending model on @huggingface, out of almost 2M open models 🚀 People sometimes forget that they’ve already transformed the field: GPT-2, released back in 2019 is HF’s most downloaded text-generation model ever, and https://x.com/ClementDelangue/status/1952827283808375168
RT @satyanadella: Excited to bring OpenAI’s gpt-oss models to Azure AI Foundry and to Windows via Foundry Local. It’s hybrid AI in action:…”” / X https://x.com/xikun_zhang_/status/1952902211278913629
Ollama and @nvidia collaborate to accelerate gpt-oss on GeForce RTX and RTX PRO GPUs. NVIDIA and Ollama are advancing their partnership to boost model performance on NVIDIA GeForce RTX and RTX PRO GPUs. This collaboration enables users on RTX-powered PCs to accurately leverage https://x.com/ollama/status/1952782326926328313
🚀We’re expanding the Tencent Hunyuan open-source LLM ecosystem with four compact models (0.5B, 1.8B, 4B, 7B)! Designed for low-power scenarios like consumer-grade GPUs, smart vehicles, smart home devices, mobile phones, and PCs, these models support cost-effective fine-tuning https://x.com/TencentHunyuan/status/1952262079051940322
RT @LangChainAI: 💻 Introducing Open SWE: An Open-Source Asynchronous Coding Agent Open SWE is a fully autonomous, cloud-based coding agent…”” / X https://x.com/Hacubu/status/1953168346356314376
RT @ori_press: We just benchmarked Qwen 3 Coder and GLM 4.5 on AlgoTune, and they manage to beat Claude Opus 4! We’re excited to see if the…”” / X https://x.com/OfirPress/status/1952470237947085146
rule number one: never distill from DeepSeek https://x.com/jxmnop/status/1953163073612562851
Is the new OpenAI open-weight model safe to release? I think the release is good for the world, but that OpenAI hasn’t ruled out substantial CBRN risks as I discuss in this post.”” / X https://x.com/RyanPGreenblatt/status/1952819470944309410
And… the biggest lesson I learned along the way? huggingface-cli uploads to a public repo by default — so add –private if you’re not ready to show your work to the world just yet. 😅”” / X https://x.com/zhuohan123/status/1952780427179258248
StepFun just released Step-3 on Hugging Face! It’s a new 321B-parameter VLM that’s “”Large yet Affordable,”” co-designed for cost-effective decoding. Achieves unprecedented efficiency, setting a new Pareto frontier for LLM inference. https://x.com/HuggingPapers/status/1952038716488208409
With the latest @huggingface accelerate release, N-D parallelism (when you stack multiple parallelism strategies on top of each other) is now a few-line configuration away. Now you can make use of one of the latest training strategies without high abstractions https://x.com/TheZachMueller/status/1953805895726489744
just one more cup before ollama gets ready https://x.com/ollama/status/1952762052755480893
LMStudio are using the upstream ggml implementation which is significantly better and well optimized. Looking at ollama’s modifications in ggml, they have too much branching in their MXFP4 kernels and the attention sinks implementation is really inefficient. Along with other”” / X https://x.com/ggerganov/status/1953088008816619637
getting ready for the day. @nvidia GeForce RTX is powered on. https://x.com/ollama/status/1952764954484027727
We’re excited to introduce a new parsing mode within LlamaCloud that lets you get complex visual recognition capabilities over documents 🖼️📑 at a cheaper price compared to pretty much anything else out there ⚡️ There’s a variety of VLM-enabled document parsing solutions out https://x.com/jerryjliu0/status/1953227974716665996
The OpenAI open weights models are very impressive. These basically beat every model from eight months ago & the small one runs on a laptop. For example, when HLE came out in January, the top score was 3-4%. Been playing with the models and so far they feel like their scores. https://x.com/emollick/status/1952796976279662596
we have a lot of new stuff for you over the next few days! something big-but-small today. and then a big upgrade later this week.”” / X https://x.com/sama/status/1952759361417466016
RT @scaling01: OpenAI has released a guide on how to finetune their open-source models: https://x.com/_lewtun/status/1952990532436934664
2.5 years later, OpenAI open source (smol) is not cracking the Dreaded Diamond Problem. https://x.com/teortaxesTex/status/1952822222298726786
OpenAI Harmony Response Format https://cookbook.openai.com/articles/openai-harmony
RT @Trinkle23897: Harmony format is finally open-sourced. I still remember 3 years ago (before ChatGPT release) @shengjia_zhao, Daniel and…”” / X https://x.com/jilin_14/status/1952800139086815641
Finally! Will be working through the internals in more detail. The first surprising fun fact is they used bias units in the attention mechanism like ye goode old GPT-2. Super interesting, I haven’t seen any other architecture doing that since then! https://x.com/rasbt/status/1952822566617501830
🚀 Introducing XBai o4：a milestone in our 4th-generation open-source technology based on parallel test time scaling！ In its medium mode, XBai o4 now fully outperforms OpenAI−o3−mini.📈 🔗Open-source weights: https://x.com/theMetaStoneAI/status/1951486506562101656
Is open weights infra on par or better than proprietary API infra now? At minimum, it feels like we’ve covered massive ground in the past few months, mostly thanks to all the infra startups listed below. If you haven’t tried open weights infra for some time, you can give it a https://x.com/ClementDelangue/status/1951668724848599143
RT @HaihaoShen: 🫡Probably the first INT4 GPT-OSS model: https://x.com/teortaxesTex/status/1953017577900228920
Spending millions to poach talent vs letting companies battle it out in the open https://x.com/fdaudens/status/1951031949184643086
The world runs on open-source. Let’s hope this continues!”” / X https://x.com/hardmaru/status/1952601984856727626
🔥BREAKING: @Zai_org’s GLM-4.5 enters the top-5 in Arena! With 4K+ community votes, it now ranks #5 Overall in the Text Arena – matching DeepSeek-R1 and Kimi-K2 as the top open models. Huge congrats to the Zai team on this incredible milestone and contribution to the open https://x.com/lmarena_ai/status/1952402506497020330
Hugging Face 🤝 Jan You can now use Hugging Face as a remote model provider in Jan. Go to Settings -> Model Providers -> add your Hugging Face API key. Then open a new chat and pick a model from @huggingface. Works with any model in Hugging Face in Jan. https://x.com/jandotai/status/1952248389531570333
RT @lexfridman: Huge thanks to all the open source projects that’ve made a lot of the tech we rely on in the world possible: Linux Git FFm…”” / X https://x.com/ClementDelangue/status/1952648566242992488
Enjoy these lil models. A lot of people worked very hard on them :)”” / X https://x.com/kaicathyc/status/1952777050298925075
2024: everyone releasing their own Chat 2025: everyone releasing their own Code”” / X https://x.com/karpathy/status/1951577221753094399
💡GPT-OSS 20B 2/4 bits GGUFs are available. Enjoy! https://x.com/HaihaoShen/status/1953729639081554002
interesting swiglu variant from the gpt-oss model: clamps inputs and adds a skip connection https://x.com/vikhyatk/status/1952808827281391701
New projects already being built on GPT OSS! Build your own with our Model APIs here -> https://x.com/basetenco/status/1952882156059148737
Next to Qwen3 of comparable size: Looks like gpt-oss is a wide (vs deep) model https://x.com/rasbt/status/1952842273848279364
One line of code is all it takes to fine-tune the gpt-oss models from @OpenAI 🔥 > Support to target the MoE expert layers with PEFT > Kernels for FlashAttention3 & MegaBlocks > Fast inference with MXFP4 quantization format In our testing, these models are extremely efficient https://x.com/_lewtun/status/1952788132908404941
openai/harmony: Renderer for the harmony response format to be used with gpt-oss https://github.com/openai/harmony
qianwen-res.oss-cn-beijing.aliyuncs.com https://qianwen-res.oss-cn-beijing.aliyuncs.com/
RT @CerebrasSystems: OpenAI GPT-OSS-120B is live on Cerebras 3,000 tokens/s – fastest OpenAI model on record 1 second reasoning time 131K c…”” / X https://x.com/cline/status/1952960760759632025
RT @mattshumer_: It’s over. OpenAI just crushed it. We have their o3-level open-source model running on @GroqInc at 500 tokens per second.…”” / X https://x.com/JonathanRoss321/status/1953119620103381440
RT @reach_vb: BOOOOM! You can now run @OpenAI gpt-oss 20B natively in @GoogleColab T4 for FREE! 🔥 Powered by Transformers ⚡ The setup tak…”” / X https://x.com/_lewtun/status/1953441199253069936
RT @thanosthinking: running gpt-oss:20b on @ollama with Turbo and web search 🏎️ 💨 very happy with how the web search turned out 🙂 and o…”” / X https://x.com/ollama/status/1952882173255856223
We’re thrilled to announce Axolotl v0.12.0. We’re ramping up our distributed training featureset with ND parallel multi-node training, and FP8 support. We’ve also added fine-tuning for gpt-oss, FSDP support for TiledMLP, and many more exciting features. 1/5 https://x.com/axolotl_ai/status/1953845149391630472
The Harmony format from gpt-oss is now supported for datasets on the @huggingface Hub 🧘 Nifty feature by @calebfahlgren! https://x.com/_lewtun/status/1953870411050959110
DeepSeek-R1: 2.66 million H800 hours GPT-OSS-120B: 2.1 million H100 hours https://x.com/scaling01/status/1952784655838564376
OpenAI claims that GPT-5 is the leading agentic tool calling model. State-of-the-art performance (97%) on the Tau benchmark. GPT-5 also achieves significant improvements in instruction following across different benchmarks. https://x.com/omarsar0/status/1953516984672420041
🚀 Qwen3-30B-A3B-2507 and Qwen3-235B-A22B-2507 now support ultra-long context—up to 1 million tokens! 🔧 Powered by: • Dual Chunk Attention (DCA) – A length extrapolation method that splits long sequences into manageable chunks while preserving global coherence. • https://x.com/Alibaba_Qwen/status/1953760230141309354
Qwen3-Coder is now available on Cerebras, 17x faster than on GPU providers. And it’s completely free. Try it out directly in your developer flow, or signup for our virtual hackathon tomorrow. It’s a $5,000 prize 🙂 @CerebrasSystems @cline https://x.com/SarahChieng/status/1951453803905163693
Small but mighty! Qwen3-Coder-Flash and GLM-4.5-Air are now on @FireworksAI_HQ Despite being smaller and faster, Qwen3 Coder Flash 30B and GLM 4.5-Air achieve almost the same quality as their larger counterparts on tool use benchmarks. The secret of good model behavior is in https://x.com/dzhulgakov/status/1952049826067050735
🚀 Meet Qwen-Image — a 20B MMDiT model for next-gen text-to-image generation. Especially strong at creating stunning graphic posters with native text. Now open-source. 🔍 Key Highlights: 🔹 SOTA text rendering — rivals GPT-4o in English, best-in-class for Chinese 🔹 In-pixel https://x.com/Alibaba_Qwen/status/1952398250121756992
💡 You get 2,000 free Qwen Code runs every day! Run this one simple command: npx @qwen-code/qwen-code@latest Hit Enter, and that’s it! 🚀 Now with Qwen OAuth support — super easy to use. Try it now and supercharge your vibe code! 💻⚡ Github： https://x.com/Alibaba_Qwen/status/1953835877555151134
Just included example scripts for aligning models using GSPO (including VLM example) 🙆♂️🙆♂️ GSPO is the latest RL alignment algo by @Alibaba_Qwen and it’s already supported in the latest TRL v0.20 release. Super-easy-to-get-started example scripts below, GO run them! 👩💻👩💻 https://x.com/SergioPaniego/status/1952305247411691871
Qwen-Image demo on Hugging Face getting absolutely hammered right now 😀 https://x.com/victormustar/status/1952416615351366033
Qwen-Image: Crafting with Native Text Rendering | Qwen https://qwenlm.github.io/blog/qwen-image/
RT @Alibaba_Qwen: 🚀 Introducing Qwen3-4B-Instruct-2507 & Qwen3-4B-Thinking-2507 — smarter, sharper, and 256K-ready! 🔹 Instruct: Boosted ge…”” / X https://x.com/NandoDF/status/1953223478087143640
RT @Alibaba_Qwen: 🚀 Meet Qwen-Image — a 20B MMDiT model for next-gen text-to-image generation. Especially strong at creating stunning graph…”” / X https://x.com/mervenoyann/status/1952455331205841261
So, I did some coding this week… – Qwen3 Coder Flash (30B-A3B) – Mixture-of-Experts setup with 128 experts, 8 active per token – In pure PyTorch (optimized for human readability) – in a standalone Jupyter notebook – Runs on a single A100 https://x.com/rasbt/status/1951635208375034191
Today we release the APIs of our Flash series, which support Qwen3-Coder and Qwen3-2507 now. Both APIs support the context length of 1M tokens. They are fast and accurate, and they are cost-effectve as well. Feel free to take a try! Qwen3-Coder-Flash Model Card:”” / X https://x.com/Alibaba_Qwen/status/1952767585596145773
@ostrisai The VAE is a fine-tune from the Wan 2.1 VAE for image generation, which is super cool and shows how open source foster collaboration, even between rival labs”” / X https://x.com/multimodalart/status/1952409238413684901
@BasedBeffJezos It’s high time we open sourced Grok 2. Will make it happen next week. We’ve just been fighting fires and burning the 4am oil nonstop for a while now.”” / X https://x.com/elonmusk/status/1952988026617119075
