Meta’s Muse Spark Open is now available in limited preview on GroqCloud, where Groq says it can generate more than 350 tokens per second per user. The open-weight model is built for agentic coding, with a 131,072-token context window and the ability to coordinate subagents and use computers across desktop, mobile and browser.
Groq Watch analysis
What happened
Groq announced the launch-day availability on 15 October. It describes Muse Spark Open as an 816-billion-parameter mixture-of-experts model, with 42 billion parameters active per token. The service supports text and image input, with text output, and is available by request in limited preview.
The model is designed for multi-step coding work, including refactoring across files, debugging and longer-running tasks. Groq says it can plan, act, observe and adjust, and writes a short statement of intent before each action so developers can follow what it is doing. Meta reports improved first-attempt accuracy and reliable tool use; Groq’s announcement does not provide benchmark results for those claims. Read Groq’s announcement.
Why it matters
Agentic coding tasks involve repeated model calls, so response speed can shape how long a developer waits for the whole job. Groq’s stated rate of more than 350 tokens per second per user gives prospective preview users a concrete performance claim to test, while the large context window is intended to accommodate substantial codebases and their dependencies.
This is a hosted preview, not a general release or a promise that every coding task will fly. The useful question is whether the speed and tool use hold up on real work, not just whether a model can sprint through a short demo.
Our read
The combination of open weights, a long context and agent-oriented tools makes this a notable addition to the coding-model field. Groq has put a specific speed figure on its hosting offer, which gives developers something measurable to challenge. Treat the preview as an invitation to test it on your own workload, not a verdict delivered by the tokens-per-second counter.
What to watch
- How developers can request access, and when the preview might broaden.
- Whether independent users reproduce Groq’s stated speed on substantial coding tasks.
- How Muse Spark Open performs on coding accuracy and tool use outside Meta’s reported results.
Discussion spark: For an AI coding agent, which matters more in practice: fast responses, a larger context window, or reliable tool use?
Sources and evidence
- Source update (15 October 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Groq. Not affiliated with or endorsed by Groq; this is not an official account.