1/ muse spark 1.2 is a very strong multimodal model—it can do visual coding, robotics planning, and audio-visual understanding that all come together through agentic tools.

2/ muse spark 1.2 performs quite strongly across a wide variety of multimodal capabilities and evals.

3/ you can get a deeper look at all this in the research blog published today.

i thought this was the oss release of muse spark :( but cool to see more focus on the multimodality!

Reading now. When can we expect watermelon Alex? And how are evals stacking up

Hopefully, Muse Video will be SOTA not on Arena benchmarks but in real use cases.

Alexandr I thought this was gonna be watermelon 🍉 and blow my mind

awesome! honered to have led the multimodal agent eval / wild artifact bench efforts

This is an incredible product. Science fiction stuff, stood up in < 18 months, absolutely incredible. So this ? Isn’t meant to patronize or belittle the achievement. aside from productizing BI tools across platforms, how will $META go to mkt to drive adoption among developers as coding and developer mkt highly saturated by OAI & Anthropic?

Alex, 1. When will muse code be open sourced? 2. Do you have an eta for spark 1.2? when that will be open sourced

Alexandr, you should check this out: I hope if you will collaborate with @humynlabs, that could bring times more accurate robotics planning ^^

Definitely one of the best for visual tasks

Ok the multimodal capabilities benchmark looks pretty good ngl! Also check out TTFT

when will muse spark support audio as input?

Muse RBD vibe coding app when?

honestly for coding.. deepseek v4 flash &gt;&gt;&gt;&gt; muse spark 1.2

one model to code them all 💀

Is anyone interested in a company that lies to society!?

My Facebook account was asked to verify its identity. After scanning my face, my account was permanently closed. What happened?

"visual coding, robotics planning, audio-visual understanding." three new adjectives for the same gpu tax

Visual coding is the piece that changes workflows fastest. Once a screenshot reliably becomes working code, the bottleneck moves from writing the UI to specifying what correct actually looks like. The part I want to see stress tested is the audio-visual side on long video. That is where most multimodal models still drift.

End-to-end models have a huge advantage in debugging with visual input providing immediate in the loop feedback on code artifact correctness

The real test isn’t better coding or planning. It’s whether multimodality + agency closes the gap between perception and consequence—or just hides it.Physical errors don’t roll back. Does stronger performance reduce residual risk… or only raise confidence in incomplete models?

Lease compute to Antrophic.

Robotics and coding in one model, that's impressive 🤖

Awesome. The next evaluation bar is whether these models can sustain plans across multi-step tasks.

"very strong" is what you call a model when the benchmark numbers can't say it for you

Drop the 🍉

I tried in open code, it’s not good to understand input.

Can we get contribution tier in europe please I want to use it for universtiy projects

Multimodal chaining through tools suggests the bottleneck shifts from model capability to how reliably agents compose outputs across modalities: where does error propagation become the limiting factor?

I wish I could edit documents on the right side of the UI directly rather than having to click further onto the artifact