Muse Spark 1.2

Meta Superintelligence Labs' multimodal reasoning model for agentic tasks, with major gains in tool and computer use, coding and multimodal understanding. Version 1.2 is the coding-optimised checkpoint, built for long-horizon multi-agentic workflows such as multi-file refactors and debugging sessions that run well past a single prompt, with a million-token window so an entire repository fits in one session. It was trained with asynchronous and parallel tool calls plus planning, goal conditioning and context compaction to hold focus across extended tasks.

Meta did not ship this checkpoint on its own. It arrived with Muse Code, a terminal coding agent, and the two were co-trained - the lab says it folded rejection-sampled harness trajectories into training along with recipe work for goals, context compaction and subagents, and integrated the agent's own toolset so the model and the harness would not disagree about how a task is run. That is the frame the release asks for: this is a model tuned to a specific way of working rather than a general upgrade.

The training emphasis is long-horizon work, and Meta names the material: whole-repository generation, large end-to-end projects and automated research. It credits three mechanisms for holding a run together over hours - planning to sequence the work, goal conditioning to keep direction, and context compaction to carry forward only what still matters. It also describes a self-improvement loop in which the previous Muse Spark generated hard coding environments and instruction-following templates, then graded candidate solutions against them, producing training data at a scale hand-authoring could not reach.

On its own charts Meta reports 82.9% on Terminal-Bench 2.1 run through Muse Code and 59.3% on DeepSWE 1.1, against 76.2% and 53.0% for the previous version measured in a lighter harness. The lab is not claiming the top of either chart: on both it places a frontier competitor above Muse Spark 1.2, and on the software-engineering set two competitors sit above it. The gain it is selling is against its own predecessor and against the cost of the tier.

The most telling result Meta published is not a benchmark at all. It set the model to optimise GPU kernels over more than a thousand tool calls and up to twenty-four hours, writing, compiling, profiling and improving Triton implementations against a reference, with third-party kernel libraries explicitly forbidden so the model had to implement the algorithms rather than wrap someone else's. It reports substantial and continuing improvement over the baseline across that window. The launch demo makes the multimodal half concrete in the same spirit: a walkthrough video handed to the terminal as a file, turned into a working booking page. Meta frames the release as a step rather than a destination, saying larger and much more capable models are on the way.