Half a million downloads in a month. Today, our open source family takes another step forward. Thank you for the incredible support behind our first-generation models. We’re excited to introduce TwIL-LM3-Pro. At just 3.6 billion parameters, it brings powerful reasoning to everyday computers, with quantized builds that run locally. No cloud required. In our evaluation: Formal logic: Highest recorded headline score among the small models compared—beating China’s VibeThinker-3B by 35% and Qwen3.5-4B by 24%, and Liquid AI’s LFM2.5-8B-A1B by 47%. Broader reasoning: 95% on SVAMP and 64.1% on MuSR, the highest recorded scores among the small models compared. BIG-Bench Hard’s logic subset: 95.4%, compared with VibeThinker-3B’s 61.1%. We believe AI is entering a post-training era. The advantage will increasingly belong to companies with the best pipelines and those that can produce capable, personalized intelligence faster and more efficiently, then put it on devices people already own. That’s what we’re building at webAI. And we’re only beginning to share what’s coming out of our lab. Coming soon: Meridian, our family of frontier-class models built to run on device. Our most advanced models will be available through the @thewebAI application. Join the waitlist as we expand access. Proudly built in Austin, Texas. 🇺🇸

Half a million downloads in a month. Today, our open source family takes another step forward. Thank you for the incredible support behind our first-generation models. We’re excited to introduce TwIL-LM3-Pro. At just 3.6 billion parameters, it brings powerful reasoning to everyday computers, with quantized builds that run locally. No cloud required. In our evaluation: Formal logic: Highest recorded headline score among the small models compared—beating China’s VibeThinker-3B by 35% and Qwen3.5-4B by 24%, and Liquid AI’s LFM2.5-8B-A1B by 47%. Broader reasoning: 95% on SVAMP and 64.1% on MuSR, the highest recorded scores among the small models compared. BIG-Bench Hard’s logic subset: 95.4%, compared with VibeThinker-3B’s 61.1%. We believe AI is entering a post-training era. The advantage will increasingly belong to companies with the best pipelines and those that can produce capable, personalized intelligence faster and more efficiently, then put it on devices people already own. That’s what we’re building at webAI. And we’re only beginning to share what’s coming out of our lab. Coming soon: Meridian, our family of frontier-class models built to run on device. Our most advanced models will be available through the @thewebAI application. Join the waitlist as we expand access. Proudly built in Austin, Texas. 🇺🇸

Viggle Turbo v0.3 for Qwen-Image-2.1 is out! - Less grain than v0.2.1, a touch softer - New 9-step mode: finer detail, small text - ComfyUI: LoRA or single-file int8/fp8/GGUF Model: https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo

We built a new way to train contextual embedding models, which encode each chunk of a document with the whole document in view. pplx-embed-v2-context-9b-preview sets a new state of the art on ConTEB and @turbopuffer's new, privately held context-bench. https://www.perplexity.ai/hub/blog/contextual-embedding-beyond-the-gold-passage

Very happy to announce that our team @GoogleDeepmind has pushed the boundaries of generative biology, achieving the successful synthesis of AI-designed proteins that are both functional and watermarked. This proof-of-concept watermarking of the building blocks of life is enabled by SynthID Bio, our new protein watermarking method. It is designed to safeguard the new era of AI-powered generative biology and strengthen global biosecurity. You can read my thoughts here on why watermarking AI-designed proteins is an important research breakthrough: https://x.com/pushmeet/status/2105314763148321102

We're joining forces with @huggingface to identify the software libraries that scientific model contributors rely on most and explore opportunities to support the maintainers behind them. https://os4science.org/news/hugging-face-open-source-for-science-fund/

getting acquired by @nvidia = hugging face can now hire people we couldn't as a small startup and give them a decade to make open-source AI win! if you're one of them, my dms are open

从 DeepSeek 官方 API 处统计的数据来看，约有 60% 的 DeepSeek Harness 用户使用了至少一个第三方插件。第三方插件是 DeepSeek Harness 用户体验中最具特色且不可缺少的一部分。DeepSeek Harness 团队将持续支持第三方插件生态的繁荣发展，并推动插件 API 趋于稳定，在将来减少和尽量避免破坏性更新。 接下来的几天我个人将每天推荐一个优质的 DSH 第三方插件，欢迎 DSH 插件作者在本 thread 下自荐。我会结合插件质量及后台实际统计到的插件使用量择优推荐。 DeepSeek Harness 团队祝大家中秋快乐阖家幸福！ （注：在用户使用官方 API 及模型时，DSH 会向官方 API 上报实际使用的插件包名和版本。此类上报不额外消耗 tokens。）

New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak , obliterating the prior record of 67.6s! This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with incredibly clever engineering and ML judgement. If a flop is low value on a particular step, skip it. Specifically: -(~8s) Sampled softmax. If a token doesn’t appear in a batch, skip its lm_head fwd/bwd some fraction of the time. -Sparse values. Only run an optimizer step for ngram embeddings that occurred in the batch. Set beta1 to zero to enable this. Beta2 is applied retroactively when the row is later used. -Sparse updates. Only update ngram and value embeddings once every 4 steps instead of once every 2. -Sparse communication. Shard the n-gram table across GPUs, and only pass the rows receiving updates on each step. -Sparse optimizer states. For the n-gram table, reduce from 2 floats in Adam optimizer per param, to 1 float per 768 params. -Hand-rolled flash attention for 64 dim heads. There are several additions that add accuracy too: -(~4s) EMA during last 300 steps, combined with lifting final_lr to 0.3 instead of 0.15. -(~1s) A new optimizer, Anvil2, which expands muon via a second tracked momentum buffer, improves the ortho coefficients, and modifies the cautious weight decay application. -A couple additional dynamic skip connections in the network. The most striking consequence of the ‘flop aware paradigm’ is you can grow parameters arbitrarily large, only limited by the available memory, since you can selectively choose how to expend flops on those parameters on each step. NanoGPT has kept active parameters below 124M, but total is unbounded, and has grown to 640M through embedding sparsity over the last year. This PR takes that to its logical conclusion on the 8xH100, scaling up to 65B sparse embedding parameters, which accounts for 25% of the PR’s gains. At frontier scale, where one is not bounded by an 8xH100, one could imagine where this paradigm could lead. https://github.com/KellerJordan/modded-nanogpt/pull/360 As this was a very notable PR, I spoke with Deven for an hour to learn how he did it. Here’s his story on the changes: https://hyperstition.cc/training-nanogpt-in-39-9-seconds

So excited to welcome @theworldlabs and @drfeifei to the @AMD family! I’ve always been a huge fan of Fei-Fei and her pioneering research in AI. Together, we’ll combine World Labs’ deep expertise in AI and world models with AMD’s compute leadership to power the future of AI and strengthen the open AI ecosystem. Can’t wait for all we’ll accomplish!