Meta opens Muse Spark Open weights: 1M-token context and 76.4 on Terminal-Bench 2.1

Muse Spark Open, aimed at agentic coding and long-horizon tasks, can now be obtained as weights and called directly on Modal. Models of this kind have generally been offered as hosted services, leaving users without weights or a way to deploy them. Meta released the weights, and Modal hosts the model as a Shared Endpoint with per-token billing.

The model combines multimodal understanding, a long context and tool use to support workflows that span many steps. Long-horizon work requires moving continuously among writing code, debugging, planning, calling tools and recovering from failure, rather than answering a single question. It has native multimodal understanding and a 1M-token context window, and supports code writing and debugging, multi-step planning, sequential tool use and failure recovery.

The published evaluation numbers place this capability on two agentic coding benchmarks. The evaluations use a named agent and harness, so the results are tied to that configuration. In Meta's evaluations it scores 76.4 on Terminal-Bench 2.1 with the Terminus 2 agent and 60 on DeepSWE 1.1 with Meta's preferred harness at xhigh reasoning effort.

The report covers the open weights, the Modal Shared Endpoint launch, and two benchmark scores from Meta's evaluations. It is aimed at engineers working on agentic coding, long context and tool use. The scores are tied to the Terminus 2 agent, Meta's preferred harness and xhigh reasoning effort; behavior under other settings requires the model card.

Both benchmark scores come from Meta's own evaluations, with settings given in the model card. The scores correspond to a specific agent, harness and reasoning effort, and performance under other configurations is not described.