Muse Spark 1.2 Guide: Meta's Coding Checkpoint and Its Harness

Muse Spark 1.2 is easiest to understand as half of a pair. Meta released it on 5 August 2026 together with Muse Code, a terminal coding agent in beta, and the two were trained together. That is useful if you plan to use Meta's agent. It is a caveat if you plan to drop the model into a harness of your own.

It has since been followed by Muse Spark 1.3, which launched in Muse Code four weeks later. 1.2 is still listed in Meta's API documentation, and there is one specific case where the docs point you to it rather than its successor.

A coding checkpoint on a general agent

Meta framed 1.2 as a coding-focused update to Muse Spark 1.1. Meta's announcement says it scaled up training compute on coding tasks while widening the variety of training environments, with gains in code generation, complex debugging, repository understanding and end-to-end developer workflows. General agent ability was meant to be kept, not traded away.

Two disclosed training details are more interesting than the headline. 1.2 was trained on whole-repository, long-horizon tasks rather than isolated snippets, and Meta describes a loop in which 1.1 generated training environments and graded candidate solutions for 1.2. Meta does not disclose a parameter count or architecture for any Spark model.

Everything else is shared across the family: a 1,048,576-token window, text output, and multimodal input. The standard price did not change from 1.1, at $1.25 input, $0.15 cached input and $4.25 output per million tokens, which is also what OpenRouter lists at the time of writing.

Co-trained with its own harness

Muse Code was co-trained with the model. That matters more than any single benchmark, because the harness decides what the model sees: how tool results are formatted, how the plan is represented, when context is compacted, how sub-agents report back.

A model trained against one harness learns its conventions. Muse Code uses persistent background agents, sub-agents in isolated git worktrees, an append-only event log that lets a session resume after a crash, and bundled /plan, /grill and /goal skills. A 1.2 session inside that loop is operating on home ground.

Put the same model in Aider, Cline or a hand-written loop and the conventions change. It may well still perform; there is simply no published evidence either way, which is why the numbers below need reading with care.

Reading the vendor numbers

The scores Meta published for 1.2 against 1.1, as summarised by Vorp Labs, are all vendor-run, pass@1 averaged over five attempts:

• Terminal-Bench 2.1 — 82.9, up from 76.2
• DeepSWE v1.1 — 59.3, up from 53.0
• Meta internal coding bench — 70.6, up from 68.3

The Terminal-Bench and DeepSWE runs used the Muse Code harness at xhigh effort. So these are the scores of the model inside the harness it was trained with, at its most expensive public reasoning setting. That is a legitimate measurement, but it is the ceiling of what you should expect, not the floor.

The internal bench is the least useful of the three because nobody outside Meta can reproduce it. Note too how small the gain is there compared with the public benchmarks; a model trained hard on coding moving two points on the vendor's own suite is a reminder that headline improvements are not uniform.

The audio exception that keeps 1.2 relevant

Meta's model documentation carries a caveat on 1.3: audio understanding is not fully supported and responses to audio requests may be degraded. The same page recommends Muse Spark 1.2 for audio input.

That makes 1.2 the version to route to when a request includes an audio file, even if everything else in your stack has moved to 1.3. The practical pattern is a small routing rule keyed on input modality rather than a wholesale choice of version. Model routing and fallbacks covers how to set that up without duplicating your whole client.

Third-party write-ups disagree on how complete audio support was in 1.2 itself, so test with your own recordings before you depend on it.

Standard 1.2 or the contributor tier

1.2 introduced a second model id, muse-spark-1.2-contributor: the same model at a much lower price, in exchange for granting Meta permission to train on your prompts and completions. The decision between the two is about data, not capability. The Muse Spark 1.2 Contributor guide covers the terms and the rate-limit catch.

The short rule: anything containing client code, credentials, personal data or material under a confidentiality agreement stays on standard muse-spark-1.2. Public, synthetic or throwaway work is where the contributor tier is worth considering.

Testing it outside Muse Code

The Meta Model API is documented as compatible with the OpenAI and Anthropic SDKs at https://api.meta.ai/v1, so pointing an existing agent at 1.2 is a configuration change. To find out how much of the harness advantage survives in yours:

1. Pick ten tasks from your repository history that needed more than a dozen tool calls.
2. Run each in Muse Code and in your own harness, same starting commit, same effort level.
3. Record completion, turns and total output tokens for both.
4. If your harness scores well below Muse Code, check how it formats tool results and errors before blaming the model.

If the gap is small, 1.2 travels well. If it is large, you have learned that the published numbers describe the pair, and that either 1.3 or a different model may suit your loop better. Either way, the number you measured applies to your work in a way the launch chart never could.