
Had the honor to present diffusion transformers at CS25, Stanford. The place is truly magical. Slides: bit.ly/dit-cs25 Recording: youtu.be/vXtapCFctTI?si=dlcE… Thanks to @stevenyfeng for making it happen!

ML at Hugging Face 🤗

Had the honor to present diffusion transformers at CS25, Stanford. The place is truly magical. Slides: bit.ly/dit-cs25 Recording: youtu.be/vXtapCFctTI?si=dlcE… Thanks to @stevenyfeng for making it happen!

. @OzzyGT wrote a guide on adding a model to Modular Diffusers without waiting for an official integration. The example is IDOL: one image to a 3D Gaussian splat, loaded with a few lines 🔥 You stay in one environment, and you get the same offloading, torch.compile, and attention backends as any other Diffusers model. This is for running and sharing a model, not a substitute for an official PR. It uses `trust_remote_code`, so only load repos you trust. This is going to be useful, so be sure to bookmark! huggingface.co/blog/OzzyGT/u…

PyBot: No weights. No training. A chatbot built with only simple rules and no AI, entirely written in 600k lines of Python code. PyBot matches Smollm with 360M parameters! How? A swarm of 1,000 DeepSeek Flash agents wrote the whole code in just a week. github.com/huggingface/pybot

i asked claude to implement trellis.2 fully in cuda c++, and it is 5.3x faster than the original pytorch implementation 🚀 surprisingly, the glb-export function was the bottleneck, not the neural net itself. moving exporting to the gpu gave more gains than anything else. - zero python runtime, all in c++ ⚡️ - cost 💵 ? ~$7 on claude and ~$60 on compute for testing on h200s through @huggingface jobs - it still gave 2.2x better perf while being bitwise identical 🎯 the code is open-sourced for use here -- github.com/suvadityamuk/trel… use it as a source for your agent to optimize your models, or utilize it directly. happy hacking! 🤗

As @ariG23498 says, “Tokyo, we are going to be in you”.
PyTorch Day Japan 2026 is just over three weeks away. We’ll be in Tokyo on December 10 for talks on deploying models on edge devices, adapting inference to different AI accelerators, and improving training performance. Michiaki Tatsubori and Tatsuhiro Chiba of IBM Research (@IBMResearch) will present “Beyond GPUs: A Practical Playbook for Bringing ExecuTorch to a New Edge Accelerator.” In the afternoon, Aritra Roy Gosthipaty (@ariG23498) and Sayak Paul (@RisingSayak) of Hugging Face (@huggingface) will present “Kernels Without Borders: Unifying the Kernel Ecosystem.” Masahiro Tanaka (@toh_tana) of the DeepSpeed team (@DeepSpeedAI) will close out the talks with “DeepSpeed at the Frontier: Community-Driven Optimizations for State-of-the-Art Training Performance” before the attendee reception. PyTorch Foundation (@PyTorch), Hugging Face (@huggingface), IBM (@IBM), and Mitsubishi Electric (@ME_JP_official) are co-hosting the event. Register for PyTorch Day Japan at events.linuxfoundation.org/p…

Massive work led by @halcyonrayes!
An end 2 end fine tuning guide for 3D model just dropped in Hugging Face. 😱 huggingface.co/blog/3d-lora-…

Cool CUDA kernel tutorial I came across today. Simon Boehm walks a matmul from naive -> coalescing -> shared mem -> tiling -> warptiling, toward cuBLAS-level FP32. Coalescing ALONE jumps ~1% -> ~8.5% of cuBLAS in his table. siboehm.com/articles/22/CUDA…

Today, it is more important than ever to fight for a future where AI is not controlled by just two companies. On this day we release the v1.0.0 of the Llama.cpp libraries. The foundation is now stable for the thousands of applications built on top of Llama.cpp and the millions of developers building them. As local AI, privacy and sovereignty will become increasingly important in our everyday lives, we reiterate our commitment to covering every platform and every model type out there, while keeping the library robust and lean. Kudos to @ggerganov, @mervenoyann, and team 🖤
llama.cpp v1.0.0 is now available huggingface.co/blog/ggml-org…

H3-Image from @MiniMax_AI is #2 on the Hub's trending models this week: 765 likes and 142,989 downloads, with a ComfyUI pack already at 351,803 downloads. Open-weight text-to-image and editing. huggingface.co/MiniMaxAI/Min…

You heard it here first.
we're not going anywhere!

Finding an analogy to be particularly very relevant. Back in 2019, Karpathy wrote: > As a result, (and this is reeaally difficult to over-emphasize) a “fast and furious” approach to training neural networks does not work and only leads to suffering. I think this theme is still very relevant in the age of agents. Just blindly trusting whatever your favorite agent spits out and not doing any auditing is a waste of our collective potential. Finding a way where the agent becomes your aid to achieve your work faster and more effectively (without causing misery) is a goal worth optimizing for. For me personally, it's sometimes about letting an agent come up with a draft PR and then iterating on top of it before requesting for any reviews. This is particularly useful for the model integration PRs where we have dedicated agent skills that are results of our solid experience of doing integrations ourselves. But for other core stuff like refactors, new modules, new features, etc. it's all a mixed bag, really. Sometimes, I ask an agent to come up with a planning doc only to reject it later because it just won't get the abstractions (yeah I am fortunate to still be in a situation like this). But the biggest learning for me so far is to always review the work before requesting for reviews from others, ensuring that the work reflects how I would have done it, basing things off my experience. x.com/badlogicgames/status/2…
the most import skill in 2026 is still being able to tell your agents "wtf is this crap, no, not like this" if you stop doing that, you're basically fucked, despite what some devfluencers would like you to think.

the most import skill in 2026 is still being able to tell your agents "wtf is this crap, no, not like this" if you stop doing that, you're basically fucked, despite what some devfluencers would like you to think.



Bonus: Qwen-Image-2.1-Lightning from @LightX2V 4 steps, 435k downloads 🧨 Which Turbo LoRA should we cover next? huggingface.co/lightx2v/Qwen…

Krea2 Turbo Distill 4-step LoRA 🎃 4 steps, distilled from @krea_ai's 8-step Krea-2-Turbo. Diffusers snippet in the model card. huggingface.co/lvladikov/Kre…