@lumalabsai prev @unslothai ex quant @statestreet @jpmorgan
Chicago, IL
Joined November 2011
- Tweets394
- Following1.3K
- Followers252
- Likes2.4K
One of the most interesting projects I’ve worked on with a cracked team. We taught a 27B VLM to control robots with GRPO, entirely in simulation.
Lots more to explore in sim-only RL.
You don't need a frontier model to control a robot. An open VLM already knows what to do; simulation can teach it how.
Learn more at zero-real.github.io
(10/11)
Mathew retweeted
Exact sequence parallelism for packed multimodal masks.
It picks ring attention, a sparse all-to-all or Ulysses per layer, matches single-GPU numerics, and beats a FlashMask + Ulysses hybrid in 47 of 48 cases on up to 128 H100s.
lumalabs.ai/news/packed-mult…
Mathew retweeted
Spent some time creating a lightweight, performant async RL stack in pure JAX with @AdityaMakkar0! We share a work log covering inference, RDMA weight transfer, memory optimizations, sharding and a few more JAX engineering insights we've seen along the way! It’s a bit of a long read, but I hope it sheds some light on the systems side of scaling RL!
We added a UI for finetuning and chat!
Introducing Unsloth Studio ✨
A new open-source web UI to train and run LLMs.
• Run models locally on Mac, Windows, Linux
• Train 500+ models 2x faster with 70% less VRAM
• Supports GGUF, vision, audio, embedding models
• Auto-create datasets from PDF, CSV, DOCX
• Self-healing tool calling and code execution
• Compare models side by side + export to GGUF
GitHub: github.com/unslothai/unsloth
Blog and Guide: unsloth.ai/docs/new/studio
Available now on Hugging Face, NVIDIA, Docker and Colab.
Mathew retweeted
Merry Christmas from Unsloth! 🎄🎁
Thank you for all the support this year!
We’re excited to keep shipping open-source next year 🥰
Mathew retweeted
Today, we launch the ElevenLabs OSS Engineers Fund - a program that provides sustained support to the open-source projects that help power our work.
Over the next six months, we are contributing $22,000 to projects our engineers rely on.
You know those tasks where you’d like to load massive context? You can now finetune on those long context tasks in Unsloth.
You can now do 500K context length fine-tuning with Unsloth!
Train gpt-oss-20b to extend its context window to 530K on 80GB VRAM & 750K+ on 192GB - no accuracy loss.
Unsloth's new algorithms + Tiled MLP = 72% less VRAM & 6x more context
Blog + Notebook: docs.unsloth.ai/new/500k-con…
Mathew retweeted
We made a free notebook that fine-tunes IBM Granite 4.0 into a powerful support agent!
This agent will enable real-time analysis & solving of customer interactions.
You'll also learn how to train models using data from Google Sheets.
Colab Notebook: colab.research.google.com/gi…
Mathew retweeted
Join us, @Pytorch and @AMD for a Virtual Hackathon on Oct 18-20. 🔥
Win $10K in prizes by training the best AI agent via Unsloth!
⭐ Sign up here: luma.com/4i64p3ec
Mathew retweeted
We're teaming up with @MistralAI and @NVIDIA for an Unsloth event on Tues, Oct 21 at @YCombinator's office! 🦥
Join us in San Francisco for a night of talks, merch and more.
Food & drinks provided. RSVP required!
⭐ lu.ma/unsloth-yc
Mathew retweeted
We managed to save >50% VRAM for multimodal RL in @UnslothAI with vLLM weight sharing! GSPO, Dr GRPO and much longer contexts are also possible!
Our notebook shows how you can design good reward functions to teach a VLM to answer hard maths and logic questions with RL
Currently Gemma 3, Qwen 2.5 VL, Llama Vision are supported & Mistral 3 + others are incoming!
Docs: docs.unsloth.ai/new/vision-r…
Colab: colab.research.google.com/gi…
You can now train Vision LLMs with Reinforcement Learning in our free notebook!
Unsloth VLM RL via GRPO: 1.5× faster, 90% less VRAM, 15× longer context & no accuracy loss.
Guide: docs.unsloth.ai/new/vision-r…
GitHub: github.com/unslothai/unsloth
Qwen2.5-VL Colab: colab.research.google.com/gi…
Mathew retweeted
The @GoogleDeepMind Gemma 3n Challenge ($150,000 in prizes) ends in 7 days!
We've made 2 new fine-tuning Gemma 3n Kaggle notebooks (Vision & Audio) to spark your creativity. Your fine-tuned model can compete for any prize.
Notebooks + Challenge Details: kaggle.com/code/danielhanche…
Mathew retweeted
We’ve teamed up with @GoogleDeepMind for a challenge with a $10,000 Unsloth prize! 🦥
Show off your best fine-tuned Gemma 3n model using Unsloth, optimized for an impactful task.
The entire hackathon has $150,000 prizes to be won!
Kaggle notebook: kaggle.com/code/danielhanche…
Mathew retweeted
We made a complete Guide on Reinforcement Learning for LLMs!
Learn about:
• RL's goal & why it's key to building intelligent AI agents
• Why o3, Claude 4 & R1 use RL
• GRPO, RLHF, DPO, reward functions
• Training your own local R1 model via Unsloth
🔗docs.unsloth.ai/basics/reinf…
Mathew retweeted
We just crossed 10 million monthly downloads on @HuggingFace! 🦥🤗
It's all thanks to you guys - the amazing community, model builders, and HF team! 💖
Something I helped with. Check it out and let me know what you think!
You can now fine-tune TTS models with Unsloth!
Train, run and save models like Sesame-CSM and OpenAI's Whisper locally with our free notebooks.
Unsloth makes TTS training 1.5x faster with 50% less VRAM.
GitHub: github.com/unslothai/unsloth
Docs & Notebooks: docs.unsloth.ai/basics/text-…
Mathew retweeted
ok in return, as curator i'm going to pick one ai.engineer 2024 video that y'all are still sleeping on
youtube.com/watch?v=pRM_P6Uf…
this is literally the @danielhanchen masterclass and I'm honestly so impressed with his depth and yet approachability/ability to explain concepts simply that others have to resort to jargon and symbols for.
very honored that he's coming back to AIE 2025 again for a part 2 on RL, because there's very few that can do what he does here at a conference of this size.
