MDL-3883EST.2026 · IDX.642
Language modelIn production

Muse Spark 1.2

Meta Superintelligence Labs · USA · 2026

Meta's coding-focused refresh of Muse Spark, trained on whole-repository work and shipped together with the Muse Code terminal agent.

wujec.ai score

8.6/10

Community score

no votes yet
Sign in to rate

Muse Spark 1.2 is the model Meta Superintelligence Labs released on 5 August 2026, four weeks after Muse Spark 1.1. Meta describes it not as a new generation but as a coding-focused update: the same family, with training compute on coding tasks scaled up substantially and the diversity of training environments widened. What the company says it trained on is more interesting than the headline. Muse Spark 1.2 was optimised for long-horizon software work — whole-repository generation, end-to-end projects and auto-research — rather than for isolated function-completion puzzles, and it leans on planning, goal conditioning and context compaction to stay coherent across those runs. Meta also states that part of its training data was produced by its own predecessor: Muse Spark 1.1 generated material used to train 1.2, a self-improvement loop the company puts in writing rather than hinting at. The model shipped alongside Muse Code, a terminal coding agent co-trained with it. The two were developed together, which is Meta's explanation for why the pairing performs better than the model alone. Muse Code is a tool rather than a model and has no separate profile in this catalogue. On benchmarks Meta claims leads on Terminal-Bench 2.1, DeepSWE 1.1 and its own Meta Internal Coding Bench; the blog publishes the comparisons as charts without a table of figures, so this profile does not restate numbers Meta did not write out. Independent measurement from Artificial Analysis places the model at 54 points on its Intelligence Index. Meta has not disclosed parameter count or architecture, and has not released weights — access is through the Meta Model API and the Muse Code beta, with what the company calls expanded global availability. Context window and API pricing carry over unchanged from Muse Spark 1.1 according to third-party reporting rather than a Meta price sheet, and are marked as such below.

#agentic coding#multimodal#long context#Meta Model API#closed weights
Official website

News

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review