Machine Learning Engineer @huggingface Open Source team
New york
Joined February 2023
- Tweets774
- Following530
- Followers1.7K
- Likes3.6K
Marc Sun retweeted
LFM2.5-24B adapters for Pi and Claude Code!
We release LoRA adapters trained on top of LFM2.5-24B - more efficient agentic development on opensource harnesses (up to 40% less tool calls on Pi!)!
huggingface.co/collections/h…
We rebuilt NVIDIA's viral speaker-captions demo with transformers serve 🎬
Nemotron 3.5 Transcribe:
→ top open-weight model on the Open ASR Leaderboard
→ promptable 30B MoE with 4B active
→ 40 languages, up to 1h of audio
→ speaker attribution, translation, audio understanding
All from one call to `/v1/audio/transcriptions` endpoint
Try it:
transformers serve nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
Model: huggingface.co/nvidia/NVIDIA…
Blogpost: huggingface.co/blog/nvidia/n…
Kudos to @eustachelb for the 0-day integration in transformers !
PR needed for the demo : github.com/huggingface/trans…
Marc Sun retweeted
TRL's AsyncGRPOTrainer tree attention now supports linear attention models 🤗. Up to 3.46x speedup!
We started from FLA’s chunked Triton implementation and changed two state-scan kernels:
• Forward: each branch inherits its parent’s final state.
• Backward: the parent sums gradients from its children.
We process the shared prefix once, then pass its final recurrent state to each continuation. Those continuations can run in parallel. In backward, their gradients are summed back into the shared prefix. We also fix the correct tokens for the causal conv1d layer.
Marc Sun retweeted
score centering + masked importance sampling is now the default in TRL's async GRPO Trainer!
Across various staleness and sampler quantization settings, the stability is pretty wild 👇
Marc Sun retweeted
We gave small, local, open models a stripped game binary. No source, no symbols, no docs.
Their job: rebuild every function in Python until the port matches the original byte for byte, on every tick.
One lead agent ran a team of small workers. The best part: when the workers kept making the same mistakes, the lead rewrote their harness mid-run.
Write-up 👉 huggingface.co/blog/ybelkada…
🤗 transformers serve now speaks 🔊
We just added an OpenAI-compatible `/v1/audio/speech` endpoint.
And we couldn't have picked a better model to launch it
with than Phonon, @GradiumAI brand-new on-device TTS: tiny, fast and multilingual.
Run with:
transformers serve gradium/phonon-40m-en
Huge kudos to @eustlb for the day-0 transformers integration, and @GradiumAI for open-sourcing their models!
Phonon-31m is Apache 2.0 🔥
huggingface.co/gradium/phono…
Marc Sun retweeted
A new frontier for on-device speech.
We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.
Phonon-31m is available under Apache 2.0.
gradium.ai/phonon🧵
Marc Sun retweeted
Harness Lens - what harness for your use-case?
Compare 19 (and counting!) harnesses across tools, design, and how opinionated they are.
Learn about how each compacts, their system prompt, and what makes them 𝙩𝙝𝙚𝙢.
Impressive tool by @yb2698: huggingface.co/spaces/ybelka…
Marc Sun retweeted
There are 2,274 open-source decision models on the Hub. Which one should you actually use?
Now there's an answer: typed-decisions is officially a benchmark on the Hub 🏆
Live leaderboard, filterable by size, from <1B to <500B. The community moves fast, and now you can see who's winning 👇
huggingface.co/datasets/Loca…
Marc Sun retweeted
Adapt your small model on your target harness with less than 20$ using RL + HF jobs !
We improved Qwen3.5-0.8B from 10% to 36% on our test set by reducing the number of tool calls by almost 2.
Everything fits on a single 24GB GPU and no need to generate synthetic trajectories !
🧵 👇
Marc Sun retweeted
There are 505 "agent traces" datasets on the @huggingface Hub (17.6 GB).
I parsed them to answer: what models, which harnesses, what tasks, how long are sessions?
Real user usage is ~15% of sessions. Here it is broken down:
🧵 👇
Marc Sun retweeted
Could not post last week as beam was released!
Now we are working on a dtype plan for arbitrary quantization per layer!
MOE fans, this one is for you 👀
Marc Sun retweeted
Tell your robot what you want, in plain words.
New in LeRobot: a vision-language model breaks your goal into skills your robot's brain already knows, steers it step by step, and answers questions as it works.
Demo: huggingface.co/spaces/lerobo…
Marc Sun retweeted
Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs have shared close to $1B worth of training compute on the Hugging Face Hub
Free to download
End of 2023 it was under $40M
Wikipedia was written in volunteer hours. This AI age commons is written in GPU-hours and data
(it's a floor btw: about half comes from official compute numbers rest is educated guess and it still miss the ~900k community fine-tunes)
Space with data to explore: thomwolf-hub-dollars.static.…
Marc Sun retweeted
Meta is releasing Muse Spark Open as the most-capable US open-weight model.. 🥁 and it's an agent focused MoE that takes all modalities 🔥
Welcoming the model to Hub with day-0 support in transformers and llama.cpp 🤗
huggingface.co/blog/meta-mus…
