Visualization of domains for which TD-MPC2 has been applied, including locomotion, manipulation, dexterous hands, humanoids, autonomous racing.

I finally joined 🦋! Some of you may recognize me from other sites. Here's a quick intro for new connections:

👋 I work on RL, world models, and generalization in decision-making. I'm perhaps most well known for my work on "TD-MPC2: Scalable, Robust World Models for Continuous Control" www.tdmpc2.com

Small models? Saturating? Where I live we don't know theses words.

New Open-source reasoning model (code, dataset, and model)!

Huginn-0125: Pretraining a Depth-Recurrent Model

Train a recurrent-depth model at scale on 4096 AMD GPUs on Frontier.

Zyphra beta releases Zonos, a highly expressive TTS model with high fidelity voice cloning.

They release both transformer and SSM-hybrid models under an Apache 2.0 license.

Physical Intelligence (π) Open Sourcing π0

They are releasing the code and weights for the π0 as part of our experimental openpi repository.

Blog: www.pi.website/blog/openpi
Repo: github.com/Physical-Int...

⭐ The first foundational model available on @LeRobotHF ⭐

Pi0 is the most advanced Vision Language Action model. It takes natural language commands as input and directly output autonomous behavior.

It was trained by @physical_int and ported to pytorch by @m_olbap
👇🧵

When it rains, it pours.

Baichuan releases Baichuan-Omni-1.5

Open-source Omni-modal Foundation Model Supporting Text, Image, Video, and Audio Inputs as Well as Text and Audio Outputs.

Both model ( huggingface.co/baichuan-inc... ) and base ( huggingface.co/baichuan-inc... ).

Latest #AI benchmark results: DeepSeek-R1 (including its distilled variants) outperforms OpenAI's o1-mini and preview models. And the Llama 3 distilled version now holds the title of the highest-performing LLM I've tested locally to date. 🚀

Tsconfig option to disallow features requiring transformations which are not supported by Node.js' --strip-types · Issue #59601 · microsoft/TypeScript 🔍 Search Terms --strip-types ✅ Viability Checklist This wouldn't be a breaking change in existing TypeScript/JavaScript code This wouldn't change the runtime behavior of existing JavaScript code Th...

TypeScript excitement 😉

Thanks to @searyanc.dev for landing the new --erasableSyntaxOnly tsconfig flag. Heading for TS 5.8 Beta next week 🎉

🔷 Guides users away from TS-only runtime features such as enum & namespace

🔷 Pairs nicely with Node's recent TypeScript support

github.com/microsoft/Ty...

Hugging Face's GRPO to TRL - the training algorithm behind DeepSeek R1

🔋Eliminates the value function from PPO to save boatloads of compute
💰 Samples N completions per prompt to compute average rewards across a group

To use it, run:

pip install git+https://github.com/huggingface/trl.git

Prime Intellect releases:

- INTELLECT-MATH, a frontier 7B parameter model for math reasoning that shows that the quality of your SFT initialization strongly impacts reinforcement learning.

Blog: www.primeintellect.ai/blog/intelle... Models: huggingface.co/PrimeIntelle...

We’ve been thrilled by the positive reception to Gemini 2.0 Flash Thinking we discussed in December.

Today we’re sharing an experimental update w/improved performance on math, science, and multimodal reasoning benchmarks 📈:
• AIME: 73.3%
• GPQA: 74.2%
• MMMU: 75.4%

SambaNova's EvaByte

The open-weight tokenizer-free language model. Their 6.5B byte-level LM—-EvaByte matches modern tokenizer-based LMs with 5x less data & 2x faster decoding!

ByteDance's UI-TARS, which can operate on your local personal device.

Project: github.com/bytedance/UI...
Desktop: github.com/bytedance/UI...
Browser: github.com/web-infra-de...
Models : huggingface.co/bytedance-re...
Paper: arxiv.org/abs/2501.12326

Introducing Kokoro.js, a new JavaScript library for running Kokoro TTS, an 82 million parameter text-to-speech model, 100% locally in the browser w/ WASM. Powered by 🤗 Transformers.js. WebGPU support coming soon!

👉 npm i kokoro-js 👈

Link to demo (+ sample code) in 🧵

DeepSeek-R1 is coming soon.

DeepSeek-R1 (Preview) Results. The model performs in the vicinity of o1-Medium providing SOTA reasoning performance on LiveCodeBench.

New sharing step on our journey towards easy-to-use fully-open models.

📢 Paper + code release 📃💻

After 2 years of work, I'm excited to announce our newest paper, MatterGen, has been published in Nature!
www.nature.com/articles/s41...

We are also releasing all the training data, model weights, model code, and evaluation code on GitHub!
github.com/microsoft/ma...

TinyBVH has been updated to 1.2.5 on main. New:
TLAS/BLAS construction and traversal, for single and double precision BVHs, and including a brand new GPU demo: See the attached real-time footage, captured at 1280x720 on an NVIDIA 2070 laptop GPU.
#RTXoff
github.com/jbikker/tiny...

InternLM v3

- Performance surpasses models like Llama3.1-8B and Qwen2.5-7B
- Capable of deep reasoning with system prompts
- Trained only on 4T high-quality tokens

huggingface.co/collections/...

Google's Titans: a new architecture with attention and a meta in-context memory that learns how to memorize at test time as presented by one of the author - @alibehrouz.bsky.social

ViTPose -- best open-source pose estimation model just landed to @hf.co transformers 🕺🏻💃🏻

🔖 Model collection: huggingface.co/collections/...

🔖 Notebook on how to use: colab.research.google.com/drive/1e8fcb...

🔖 Try it here: huggingface.co/spaces/hysts...

Goodbye WinterCG, welcome WinterTC WinterCG, the Web Interoperable Runtimes Community Group is moving to ECMA as TC55 to be able to publish standards.

Deno is committed to web standards - that's why we co-founded WinterCG two years ago. Today marks the next step in that journey: WinterCG moves to Ecma International as technical comittee 55 (TC55).

Goodbye WinterCG, welcome WinterTC!

deno.com/blog/wintertc

Screenshot of the dataset on the Hugging Face Hub

🔍 Massive human feedback dataset for text-to-image models from RapidData
- 1.5M human responses from 152K participants
- Evaluates image coherence, style & prompt alignment
- Includes detailed error heatmaps
- Covers DALL-E, Midjourney, Imagen outputs
Available on @hf.co

ByteDance just dropped SA2VA: a new family of vision LMs combining Qwen2VL/InternVL and SAM2 with MIT license 💗

The models are capable of tasks involving vision-language understanding and visual referrals (referring segmentation) both for images and videos ⏯️

microsoft/phi-4

phi-4 is a state-of-the-art open model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets.

huggingface.co/microsoft/ph...

Thrilled to share the latest work from our team at
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥

📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act

0/9 🧵