Pinned Tweet
Today we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.
Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.
AI at Meta retweeted
muse ember, fast, cheap, and plenty performant
We recently brought together researchers from across academia, national labs, healthcare, and policy for the Open Source AI for Science Summit to explore how open models from Meta, such as Muse Glimmer, SAM, DINO v3, TRIBE v2 and UMA, are driving real-world advancements.
A couple highlights:
→ Medical Research: The University of Pittsburgh presented an assistive-robotics system built on DINOv3 and SAM 3. The team plans to demonstrate a robot that can help a person eat and is experimenting with letting users give instructions in their own words.
→ Chemistry: With the University of Chicago, we showed how UMA can simulate entire enzymes, atom by atom, potentially helping scientists understand how they work and engineer better ones for medicine, sustainability and energy.
Read the full recap: go.meta.me/4Cg6mBI
Legal work is a tough test for AI.
Muse Wonder earned the highest overall quality score in Legora’s November BAR benchmark for legal work. A strong showing on the kinds of complex tasks businesses need help with.
The November BAR results are now available, showing how AI models compare on real-world legal tasks in Legora.
A few highlights from this month’s results:
→ Muse Wonder ranks highest on overall quality and leads across four task types: analysis and answering, extraction, drafting, and editing.
→ GPT-6.2 Astra leads on comparison and review tasks, with overall quality 4% above November’s model average and cost per case 20% below average.
→ GPT-6.2 Sol delivers above-average quality at 78% below the average cost per case across models benchmarked in November.
The race among efficient models built for real work has a new leader: Muse Ember by @AIatMeta.
Ember (Max) leads on every dimension against GPT-6 Luna (Max) from @OpenAI and Claude Haiku 5.5 (High) by @AnthropicAI:
- Speed: 27s per task vs 5.8 min and 1.9 min
- Performance: +2.3% Net Improvement (Agent Arena score) vs +1.4% and +0.4%
- Cost: 1.9¢ per task vs 7.0¢ and 6.9¢
That’s ~13x faster than Luna and less than a third of the cost of either. Below find more analysis on Ember’s steerability score across Work, and a look at what real-world tasks users are shipping with Muse Ember in the Agent Arena.
AI at Meta retweeted
Meta Muse Ember is lit 🔥 (pun intended). My personal project to help me learn Japanese that would have cost ~$1K USD is now only ~$100 USD in token costs.
I would not be able to afford this project with OpenAI, Gemini or Claude.
Fable 5.2 demos look cool, but I've seen frontier models struggle with tasks that smaller models handle surprisingly well.
My own benchmarks keep reinforcing this: bigger models don't necessarily produce better results.
Now could I get this even cheaper my training my own model? I guess I'll find that out next.
AI at Meta retweeted
Japanese Tea Garden Bench - Muse Ember reachablegames.com/teagarden…
Huge step up for this offering, although hard to know what to compare it to. Maybe an upgraded Muse Spark Open? Cheap smarts are good smarts!
AI at Meta retweeted
Muse Ember surprised me with the output 👀
A lightweight model created this 3D scene using the /design feature from Command Code.
Muse Ember nailed it in one shot, while GPT-6-Luna needed two prompts to get it perfect.
Today we're also bringing Muse Ember to @GoogleCloud along with Muse Wonder and Muse Spark 1.3.
Today we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.
Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.
Pi Dash + Muse Ember == Coding beast! 🔥
Pi Dash is honored to get early access to Muse Ember over the past weekend!
🤖 10 Muse Code agents with Muse Ember working together under Pi Dash's Coordination to evolve itself:⚡ Raw SQL → SeaORM, ~178,000 lines of production code rewrite in a day
Watch it in action 👇 pidash.airepublic.com/
AI at Meta retweeted
Ready to get started with Muse Ember? Read the dev blog for details on the model, how to access, and recipes: bit.ly/4k0YNrX
Muse Ember, our smallest, most efficient model, is now available in Muse Code and on Meta Model API.
Built for agentic execution at scale, Muse Ember excels at coding subtasks, tool calls and document analysis, with support for audio and video inputs alongside text and images for your multimodal workflows.
Pair it with Muse Wonder or Muse Spark to combine efficient execution with deeper planning, reasoning and open-ended work.
We’re excited to continue building out the Muse series of models to give developers more choice over how they tackle their toughest challenges.
s
Start building: bit.ly/4z7XqLS
AI at Meta retweeted
Our first version of Muse Ember 🔥
Ember aims to maximize intelligence per GW of inference to help bring personal superintelligence to everyone.
We would love all the feedback to keep hill-climbing. Please send it our way. We will keep hill-climbing on efficiency, user experience, and better approach to train with internal frontier!
More to come 🫡
AI at Meta retweeted
Muse Ember did this in under 2 minutes on its highest reasoning effort, this model might SLAP
AI at Meta retweeted
Muse Ember 🔥 our smallest, most efficient model built for work that agents repeat all day.
It sits on the efficiency frontier, and is designed for complex reasoning execution and work at scale: tool calls, coding subtasks, routing, extraction, document parsing.
Read more here: research.meta.ai/blog/muse-e…
AI at Meta retweeted
1/ announcing our smallest, most efficient little model is here: muse ember 🔥
built for work that agents repeat all day, it gets you more intelligence per dollar than Haiku, Luna, and all other flash models at any reasoning setting and does it at a third of the cost or less!
For agents that perform hundreds of small tasks, efficiency matters at every step. Muse Ember is built to deliver more intelligence for the cost.
On the @ArtificialAnlys Intelligence Index, it achieves performance comparable to Claude Haiku 5.5 at roughly one-third of the cost per task, while scoring higher than GPT-6 Luna across all of Luna's reasoning settings.
Muse Ember is available today on the Meta Model API, at $0.10 per million input tokens and $0.50 per million output tokens.
Learn more on our research blog: go.meta.me/muse-ember
AI at Meta retweeted
Releasing Muse Ember today -- our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Available today on the Meta Model API and in Muse Code.






























