Research Scientist at Meta FAIR | Coding & Computer use agents | CS PhD @UMich
Seattle, WA
Joined May 2016
- Tweets158
- Following453
- Followers968
- Likes832
Shengyi Qian retweeted
We trained an imaginative agent! ✨
Agima: an omni agent that plans by imagining futures.
𝗦𝗙𝗧: Directly generalizes to harder tasks! 𝟮%→𝟰𝟬%, beats GPT-5.6 Sol (𝟮𝟬%).
𝗥𝗟: New ways to imagine emerge! Imagines a second plan only if the first one fails (𝟯𝟮%→𝟭𝟬𝟬%).
Shengyi Qian retweeted
✨Happy to share the release of technical report for Muse Glimmer, an open-weights 30B dense LLM that we built with emphasis on local agentic use. Dig in for plots and numbers on:
🪜 Architecture and scaling ladders
📊 Scorecard across 39 (!) benchmarks
🌍 Multilingual evaluations with Muse Glimmer supporting 137 language varieties
🧱 Pretraining and mid-training, including checkpoint averaging
🎯 Post-training: SFT, on-policy distillation, and multi-objective RL
🛡️ Privacy and prompt injection robustness baked into post-training
🏭 Synthetic data generation pipeline
⚡ Inference optimizations: quantization and speculative decoding
Looking forward to the discussions and implementations this report sparks 🚀
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
This is a very nice and comprehensive summary of muse glimmer training process written by Tim! 🐮
A few months ago we set ourselves the goal of making an open model for the claw: good for agentic, fits in a 32GB gpu. Here's what went into Muse Glimmer: 1/?
xitter.cf/AIatMeta/status/211226…
Shengyi Qian retweeted
Banger paper from Meta Superintelligence Labs on on-device agent models.
(bookmark it)
If you build agents that need to run locally, this one is worth your time.
Muse Glimmer is an open 30B model, distilled from Meta's frontier model Muse Spark and designed to fit in 32 GB of VRAM. With 4-bit quantization and a speculative decoding head, it runs on a single RTX 5090.
On MCPAtlas, it scores 75.5, against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. On SWE-Bench Pro, it scores 51.2, against 36.9 and 50.2.
Qwen3.6-27B still leads on OSWorld-Verified and TerminalBench 2.1, so the gains are concentrated in tool use and long-horizon tasks.
The model takes text and images, covers more than a hundred languages, and is released under Apache 2.0.
arxiv.org/abs/2610.24628
Chat with Paper: academy.dair.ai/papers/muse-…
Shengyi Qian retweeted
The Muse Glimmer technical report is out: arxiv.org/pdf/2610.24628v1Mu…
Muse Glimmer is our 30B open-weight model (Apache 2.0), designed for agentic workloads entirely on consumer hardware.
Looking forward to seeing how the community uses our model!
Shengyi Qian retweeted
Metas Muse Glimmer tech report is pretty interesting.
Metas Muse Glimmer report describes training agents inside simulated digital lives, complete with 10–14 contacts, weeks of messages, calendars and documents. The agents practice tasks spanning multiple days, while training trajectories containing hallucinated results or unverified claims of success are discarded.
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
Meta's Muse Glimmer report cites our 2023 weight averaging paper (COLM'24). Their final training recipe also uses uniform averaging of 600 steps apart checkpoints, (what we proposed).
Great to see our work being adopted. Thanks, @AIatMeta!
If you wanna understand the science of weight averaging keep reading 🧵 (1/n).
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
Technical report of Muse Glimmer, a small and nimble model that is well-rounded. Enjoy!
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
The Muse Glimmer technical report is out today! It is our 30B open-weight model (Apache 2.0) distilled from Muse Spark, built to run agentic workloads entirely on consumer hardware — one RTX 5090, 32GB.
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
We just released details about our work on Muse Glimmer! Check it out!
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
Very excited to share the tech report on Muse Glimmer.
It was great fun to work on this with folks at @AIatMeta.
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Shengyi Qian retweeted
Excited to share the Muse Glimmer technical report! It was a great collaborative effort across multiple teams, and I'm glad to have been part of it!
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
🚀 Honored to be a core contributor to Muse Glimmer! The model is open — and so is the homework. Our tech report lays out how we built the open-weight model from pretraining to alignment.
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Excited to see Muse Spark Open released! Three things I particularly appreciate: (1) efficiency before scaling, (2) opening up frontier architecture ideas, and (3) multimodal-native design from day one. A few personal thoughts 🧵
2/ I love seeing frontier architecture details shared openly, e.g. LatentMoE. @Kimi_Moonshot explores a similar idea with different routing and expert choices. Open weights are also about sharing the design tradeoffs that help move the whole field forward.
3/ Multimodal-native from day one matters to me. We considered multimodality from pretraining, not as a late add-on. Unlike @Alibaba_Qwen's text-only open-weight release, Muse Spark Open shares multimodal capabilities too. The foundation—and what we open—both matter.
Shengyi Qian retweeted
Muse Wonder, aka Watermelon is out today!! The consumer use cases of CUA are obvious from Muse, a particular use case I am super bullish about is verification.
CUA offers a feature-complete harness (GUI + code) for the agent to verify its creation, give itself feedback, and self-improve, aka CUA as verification.
This demo uses Muse Wonder to generate a globe web-app from a reference image and prompt. Two agents, same model: (a) a generation main agent, (b) a CUA verification sub-agent. They iterate until the CUA verifier is satisfied. This is an interesting test-time scaling with CUA.
As usual, work done with the incredible CUA and visual coding team at Meta MSL.
@taoyds, @yuning_pro, @EdwardSun0909
