(2/6) main-horse.github.io/hnet/eng-1gpu/

Ready to meet the biggest, brainiest guy in the Qwen3 family?

FineVision is not only bigger and more diverse than 3 popular open-source alternatives, models trained on it also perform significantly better. Check out all the details in the Blog Post: huggingface.co/spaces/Hugging…

Introducing EmbeddingGemma🎉 🔥With only 308M params, this is the top open model under 500M 🌏Trained on 100+ languages 🪆Flexible embeddings (768 to 128 dims) with Matryoshka 🤗Works with your favorite open tools 🤏Runs with as little as 200MB developers.googleblog.com/en/introducing…

Meta FAIR just unveiled VLWM on Hugging Face! This Vision Language World Model is a new foundation for planning with reasoning directly on natural videos. It combines reactive (system-1) and reflective (system-2) planning for SoTA performance!

This is how you create the best open dataset for VLM

By popular request: you can now branch conversations in ChatGPT, letting you more easily explore different directions without losing your original thread. Available now to logged-in users on web.

Thats looks crazy directly streamlined as mod hello ??

3... 2... 1... lift off huggingface.co/moonshotai/Kim…

Qwens new big baby is out as it seems, Qwen3 Max no information about model size yet openrouter.ai/qwen/qwen3-max

Qwen 3 Max has no "thinking", interesting

Big news: Introducing Qwen3-Max-Preview (Instruct) — our biggest model yet, with over 1 trillion parameters! 🚀 Now available via Qwen Chat & Alibaba Cloud API. Benchmarks show it beats our previous best, Qwen3-235B-A22B-2507. Internal tests + early user feedback confirm:

Lean 4 Theorem Proving Multi-turn formal theorem proving in Lean 4, where models alternate between reasoning, sketching proof code, receiving feedback. Ideal for search guided rl, process rewards, and curriculum design. By Sinatras app.primeintellect.ai/dashboard/envi…

> we've hit a data wall > pretraining is dead Is it? Today we are releasing 📄 FinePDFs: 3T tokens of new text data for pre-training that until now had been locked away inside PDFs. It is the largest permissively licensed corpus sourced exclusively from PDFs.

Freshly curated open dataset with 3T multilingual tokens from PDFs. > containing about 3 trillion tokens across 475 million documents in 1,733 languages. > new source of data, with a knowledge cutoff in february 2025 > sota performance when mixed with fineweb-edu/dclm.

Sure i will check back in couple minutes... wait why did you hangup on me

oh my god this is just crazy folks

Test it for free, i got some impressive eval results for its size on couple domain specific tasks past week. Will revisit it with SFT in near future.

Official reminder that your favorite coding asistant will be quantized to 1.58 bits in ~10 minutes have fun with smart model while its lasts, until next time cheers

one weird trick to transpose your matrix really fast in hardware but you can’t transpose it back as fast