Meta Muse Spark Open is live on Together AI

Together AI says Meta’s Muse Spark Open is now live on its platform, and the hook is hard to miss: a 1M-token context window on both serverless and dedicated infrastructure. The company is positioning it as the most capable US open-weight model to date, and as a launch that’s ready for day one rather than a vague coming-soon promise.

The model is aimed squarely at long-running agentic coding work. That means multi-file refactors, debugging sessions, and jobs that don’t fit neatly into a tiny prompt box. Together AI says the model was trained on real developer work, and says that shows up in higher first-attempt accuracy, more reliable tool calling, and better understanding of the codebases developers actually touch.

The big practical selling point is scale. A 1M-token window lets Muse Spark Open keep entire projects in context, including older code, dependencies, and thousands of files, without chopping the job into pieces. Together AI also highlights longer-horizon autonomy: the model can plan, act, observe, self-correct, and coordinate subagents when work can be split up.

There’s also a broader multimodal angle here. Together AI says the model can handle computer use across desktop, mobile, and browser, browse the web with multimodal browsing, and take in text, images, audio, and audio-video input. It also explains what it is doing before it acts, which should make long runs easier to audit. That’s a small detail with a big effect; agent systems get a lot less magical once they have to narrate themselves.

The early access list includes Factory and Ollama, both of which tested the model before launch and say they’re offering it to users from day 0. Together AI is also publishing pricing: $1.25 per 1M input tokens, $4.25 per 1M output tokens, and $0.15 per 1M cached input tokens. Developers can try it in the playground or call it through the Together API.