Research @Meta Superintelligence Labs · Lead Multimodal Agent Eval for Muse Spark · @MIT PhD · Opinions are my own NOT from affiliation
Bay Area, USA
- Tweets 824
- Following 561
- Followers 5,933
- Likes 572
Zengyi Qin retweeted
Muse Spark 1.3 ranks #2 on Agents' Last Exam (ALE) leaderboard, ranked #1 at the time of its release.
Kudos to the team! Somehow this didn't get enough buzz.
Leaderboard: agents-last-exam.org/leaderb…
5
7
56
67,638
Zengyi Qin retweeted
Introducing OM-1, our first robot foundation model, zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids.
- learned directly from human manipulation data
- no teleop/robot data
- close to human-level dexterity and efficiency
- multi-robot collab
154
479
3,499
1,745,052
Zengyi Qin retweeted
Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API. Developers can build with frontier performance without the frontier prices.
We thought showing would be better than telling, and encouraged our friends in Meta Superintelligence Labs to come up with a few demos.
One-shot prompt: Build a rendering engine from scratch in C that writes pixels directly, scales from simple primitives to complex scenes, and culminates in a striking lighting demo.
Check out the rest 👇🏻(1/4)
39
56
780
50,422
RUMOR: OpenAI Astra model used "Looped Transformers"
We studied the 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐋𝐚𝐰 𝐟𝐨𝐫 𝐋𝐨𝐨𝐩𝐞𝐝 𝐓𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫𝐬: arXiv 2506.18233
The main discovery is shown in the figure:
1. Loop does NOT increase knowledge capacity
2. Loop is effective in increasing reasoning capability
That being said, the benchmark gains does not come from knowledge gain, but from reasoning gain, e.g., the way the knowledge is being used.
Knowledge capacity and reasoning capability is disentangled by looping layers, as shown in the green dots. As model size grows (dot size grows), the knowledge capacity (x-axis) does not grow, but the reasoning capability (y-axis) grows.
In non-looped transformers, as shown in the blue dots, the knowledge (x-axis) and reasoning (y-axis) grow at the same time.
This might reveal another scaling axis that brings a new paradigm for LLM pre-training.
I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.
OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. We deeply care about this technique, as it can give us a view into how model alignment generalizes from its training distribution. I do think it is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it's a core goal of our current research program.
18
110
975
89,655
Paper link arxiv.org/abs/2506.18233
1
1
12
2,995
Zengyi Qin retweeted
Over the past few months, we’ve been thinking a lot about what it would actually take to build agents that continuously improve from their own experience. Today, we’re open-sourcing our continual learning infra, Reef.
The idea is simple: instead of treating inference as the end of the pipeline, Reef turns live agent interactions into a continuous learning loop. It serves real applications, captures trajectories and feedback as structured experience, and lets different learning recipes use that experience to improve the system.
What evolves isn’t just the model. Reef is designed to evolve the whole agent — model weights and the harness — then evaluate, version, and safely deploy those updates back into serving.
Really excited to finally share Reef we’ve been building toward continual self-improvement!
Come and check it out: github.com/Human-Agent-Socie…
And join the Discord group for more updates: discord.com/invite/5y8e5f937…
2
15
72
59,023
Zengyi Qin retweeted
Muse Voice Transcribe is MSL's first real-time audio perception model -- rolling out today. SOTA in streaming speech-to-text, it handles speaker diarization, and endpointing natively in a single model.
393
476
5,473
1,105,203
Hello everyone! 👋🕶️
I put Meta Muse Code + the Unity CLI to the test with a real Unity project to see how far I could push this agentic workflow. Instead of just generating code, I wanted to see how it could actually interact with Unity, run tests, validate changes, port the game to other platforms, and help convert an existing project into VR. 🤯
📌 Full video: piped.video/5iOIck1AWAg
Starting with a standalone Mini Golf game, I used Muse Code to:
- Control & interact with Unity through the Unity CLI
- Create Play Mode tests + run them with unity run_tests
- Build a website that stores test history, screenshots, results & execution times
- Port the game to Three.js
- Convert the original standalone Unity game to VR with the Unity CLI + Meta VR CLI
- Test the VR version with Meta XR Simulator
- Create more advanced shaders, such as procedurally generated stars, comets, and reveal/occluder effects
Let me know if you’ve any questions. I had so much testing Muse Code powered by Muse Spark 1.2 & creating various prototypes!
10
13
75
9,280
same money gets 10x work done than claude code
1/ muse code is out of beta. launching with an sdk in developer preview to build your own agents on top of it and rolling out monthly subscription plans.
copy-paste into your terminal:
curl -fsSL dev.meta.ai/install.sh | bash
5
2
51
7,765
Zengyi Qin retweeted
We checkpoint training runs without a second thought. Almost nobody checkpoints agent runs, because agents inherited an execution model built for requests that finish in seconds.
If your agent has been coding or researching for four hours and dies in the last minute, you will pay for those hours twice. And you will lose everything it learned along the way: the paths it tried, the branches it ruled out, the state it built.
Once execution state can be saved, it can also be forked. Explore three approaches from the same known good moment, keep the one that works, throw away the rest. Graph search in production become viable.
Execution checkpoints are live in Runta today.
The reason you don't let your agent near production is nothing it does can be taken back.
We shipped Execution Checkpoint: Before a migration, a schema change, a bulk write, the agent can mark a recoverable boundary. When something breaks, it resumes from the last known good moment with its full reasoning and memory intact.
Resumability is a runtime property. It has to be built into the execution layer.
3
4
20
4,431
We built the first benchmark for AI capability in marketing. Turns out "use the best model" is a coding answer, not a marketing one.
Kimi K3 (@Kimi_Moonshot) leads at 94.8, at $1.42 a task.
Ox Alpha aka GLM-5.3-Flash from @Zai_org showed up this week and went straight to #2, 89.9 at $0.87. Opus 5 is 11th of 15, at $8.43.
One real Google Ads account, six real marketing tasks, up to 30 runs per model. Every run in its own container on @modal_labs. This is AdsBench (check out adsbench.ai) 🧵👇
8
10
39
3,721
Zengyi Qin retweeted
You asked us to bring it everywhere, so we did.
Muse Spark 1.2 contributor tier is expanding globally.
Same model, same 1M context and multimodal input, at our lowest prices: $0.10 input / $0.20 output per million tokens. Let your prompts and completions help improve future Meta AI models, and you get our lowest rates in return.
Refactor, prototype and ship without watching the meter.
It’s now available in Muse Code, Meta Model API.
Start building now → bit.ly/4xTqMxs
Meet Muse Spark 1.2 and Muse Code, the agent built to run it against your coding challenges. Now in beta.
Together they take a task from idea to production, and deploy teams of agents that plan, write and refine across your repos.
Start building now: bit.ly/4fTXMy6
42
43
761
93,959
