Muse Spark 1.2: Meta's Coding-Focused Muse Release, With Muse Code

What Muse Spark 1.2 is

Muse Spark 1.2 was released on 5 August 2026. At launch it was proprietary and available only through Meta's API and Muse Code; on 20 August 2026 Meta published new evaluation results "ahead of the open-weights release" of the model, without giving a date. It shipped alongside Muse Code, Meta's first terminal coding agent. It is a reasoning model tuned for multi-agent workflows, with a 1 million token context window, configurable reasoning effort, and input across text, image, video, audio and PDF. There is no chat application for it and no published free tier — this is a developer product accessed through the API and Muse Code. Meta's 20 August post also reports that tool access raised one visual benchmark score from 59.8% to 72.0%, and describes real-time search combined with web development. Meta publish an installation command for Muse Code on macOS and Linux, and list asynchronous background agents among its features, with Muse Spark 1.2 as the model underneath, in beta at launch.

LMArena and independent figures

On LMArena's text leaderboard Muse Spark 1.2 has 1492 Elo (rank 13, xhigh effort); by category, 1543 Elo in Coding (rank 10, xhigh effort), 1492 Elo in Math (rank 28, xhigh effort), 1493 Elo in Expert (rank 55, xhigh effort) and 1441 Elo in Creative Writing (rank 58, xhigh effort). SimpleBench records 75, with SimpleQA Verified at 60.

Pricing and what is not published

Input is $1.25 per million tokens and output $4.25, with cache hits at $0.15 per million — first-party vendor pricing, unlike the later 1.3 release where Meta stopped publishing rates on the launch post. Not published: no maximum output token count, no batch discount, no code execution support, no long-context surcharge policy and no cloud marketplace availability. Fine-tuning is not offered. Web search, function calling and structured output are supported.

What Meta changed from 1.1

Meta describe Muse Spark 1.2 as a coding-focused update to Muse Spark 1.1, improving code generation, complex debugging, codebase understanding and end-to-end developer workflows. They state they significantly scaled up training compute on coding tasks while widening the diversity of training environments, co-trained the model with Muse Code using rejection sampling, and trained it extensively on long-horizon tasks such as whole-repository generation, large projects and automated research. Meta also describe a self-improvement loop in which Muse Spark 1.1 generated training data for 1.2. In a kernel-optimisation case study, the model ran for more than 1,000 tool calls over up to 24 hours on KDA and MLA kernels for NVIDIA Hopper GPUs.

Muse Code and multimodal results

Muse Code, released in beta on the same day, is a terminal coding agent that plans changes, writes code and validates results across large repositories, with asynchronous background agents, a local event log for exact replay and restart-safe runs, and bundled /plan, /grill and /goal skills. In its 20 August post on the model's multimodal abilities, Meta report ZeroBench 54.0%, SimpleVQA 75.0% and CharXiv Reasoning 87.6%, and describe visual coding of web pages and games, robotics with bimanual manipulation, and audio-visual video understanding.

Official Resources