Muse Spark 1.2

A coding-focused successor to Meta Superintelligence Labs' (led by Alexandr Wang) Muse Spark 1.1, released August 5, 2026 alongside Muse Code — MSL's first terminal coding agent (beta), which Muse Spark 1.2 powers. Meta "significantly scaled up training compute on coding tasks while expanding training environment diversity," co-trained the model with Muse Code, and trained it on long-horizon agentic software engineering: understanding large repositories, coordinating multiple sub-agents, and iterating until a task's objective is met. Meta's third model in four months, positioning the Muse line against Anthropic's Claude Code and OpenAI's Codex.

Meta reports coding gains over Muse Spark 1.1 — Terminal-Bench 2.1 82.9% (up from 76.2%) and DeepSWE 1.1 59.3% (up from 53.0%), plus a Meta Internal Coding Bench. On the broader composite it reaches AA Intelligence Index 57 (v4.1.1, xhigh) — Meta's highest AA score to date, up from 1.1's 53. 1M-token context, text and image input.

Proprietary (parameters undisclosed), available via the Meta Model API and OpenRouter at $1.25 / $4.25 per 1M input/output tokens ($0.15 cache hits), with a feedback-gated Contributor Tier at $0.10 / $0.20 (~12× cheaper). Muse Code installs on macOS and Linux with a single command.