Meta opens the weights of Muse Spark Open: agentic coding and 1M tokens already on Modal

Open weights for agentic coding are a category where developers first ask not "how smart is the model?" but "where can I run it and how much will it cost?" Modal answered the second question quickly: as soon as Meta released the weights of Muse Spark Open, the platform deployed the model on its Auto Endpoints.

What the model is and what it is for

Muse Spark Open is built for agentic scenarios. It is not about "answer a question about code" but "take a task and see it through to the end." According to the description, the model can write and debug code, build multi-step plans, call tools in sequence, and recover from failures in long chains of actions.

That last point is the most down-to-earth. Anyone who has run an agent on a real repository knows it rarely breaks at the first step. More often it happens at step twenty, when a terminal command failed and the agent carries on as if everything were fine. The ability to notice an error and roll back is sometimes more important than beautiful code on the first try.

There are two more parameters. The first is native understanding of multimodal data: a screenshot of an error or an interface mockup can be given to the model directly, without an intermediate description. The second is a context of 1M tokens. For an agent, this means that a sizable part of a repository, the history of previous steps, and logs fit in the window, and you don't need to immediately build a complex memory compression scheme. But "fits" does not mean "uses well," and that has to be tested separately.

How to read the benchmarks

Meta cites two results:

• 76.4 on Terminal-Bench 2.1 with the Terminus 2 agent;
• 60 on DeepSWE 1.1 with the harness Meta itself prefers.

Both measurements were made at the xhigh reasoning level, that is, at the highest "thinking" effort. This is an important caveat: such a mode usually consumes more tokens and time than the standard one. Terminal-Bench is about an agent's work in the terminal, DeepSWE is about software engineering. Both tests are close to what agents actually need to do.

Still, the numbers were obtained by Meta itself on settings it chose. The full evaluation conditions are in the published model card. The harness often shifts the result more than the difference between neighboring models. So comparing these numbers with others' without adjusting for the harness and reasoning mode is not valid.

How to try it and what to check

On Modal, the model runs as a Shared Endpoint: a shared endpoint with pay-per-token pricing, without renting dedicated GPUs. For a team, this is a convenient way to run a short pilot. You plug the model into your own agent loop and see what happens on live tasks.

What is worth measuring:

• the share of tasks completed end to end on your repository versus your current model;
• behavior after a tool failure: did the model notice the error and correct itself;
• the cost of a whole task, not a single token: an agent with a high reasoning level can spend a great many of them;
• quality on long context, not only on short prompts.

If the results are satisfactory and request volume grows, you can think about your own deployment. That is the point of open weights: control over data, a pinned model version with no surprises from the provider, and the theoretical possibility of fine-tuning.

What we don't know yet

The announcement does not reveal the model's size, hardware requirements, price per million tokens, or license terms. All of this should be checked on the model card and the Modal page before building it into a product. Moreover, this is a message from an inference provider, not an independent test. Until there are third-party measurements across different tasks, Meta's numbers should be treated as a claim, not as proof.

The sensible position right now is simple: Muse Spark Open is a good candidate for evaluation for those building coding agents who want open weights. Test it on your own repository, calculate the price of a whole task, and only then decide whether to switch.