Observed facts: Groq announced on October 15, 2026, that Meta’s Muse Spark Open entered limited preview on GroqCloud. Groq lists 816 billion total parameters, 42 billion active per token, and a 131,072-token context window on the platform. It also reports more than 350 tokens per second per user and says the model accepts text and image input with text output.
Mechanism: The announcement frames coding agents as sequences of dependent model calls, so serving latency can affect the time to complete a multi-step task. A hosted preview gives selected builders a way to evaluate the model within GroqCloud rather than treating an open-weight release as an abstract artifact. Why now: The deployment update arrives alongside the model’s stated launch, making access on an inference platform part of the release story rather than a later integration detail.
Implications: Builders can assess model behavior and serving characteristics in the same operational environment. This does not establish that the reported throughput will translate into faster completed work for every workload, nor that the model outperforms alternatives. Evidence boundary: The figures and capability descriptions above are Groq’s claims in its partner announcement; no independent benchmark or broad availability is established by the supplied source.
Watch condition: Watch whether Groq expands access beyond limited preview and publishes workload-specific measurements that connect serving speed to end-to-end coding-task completion.