Meta Releases Muse Code Beta With Muse Spark 1.2
Muse Spark 1.2's rapid benchmark gains, from 43 to 54 on the Artificial Analysis Intelligence Index in four months, place Meta in a near-tie for third among American AI companies, signaling a competitive shift in the AI coding agent market.
Key Facts
- Meta released the beta version of Muse Code, a terminal coding agent, on August 5, 2026.
- Muse Code runs on macOS and Linux and includes an asynchronous background agent for parallel information gathering.
- Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, ranking second after Claude Opus 5 and Claude Code's 86.7%.
- Muse Spark 1.2 scored 59.3% on DeepSWE 1.1, ranking third after Claude Opus 5's 65.0% and GPT-5.6 Terra's 64.8%.
- Muse Spark 1.2 is available via the Meta Model API and is integrated into Muse Code.
Reporting from 1 source: GIGAZINE.
Meta announced on August 5, 2026, the beta release of Muse Code, a terminal-based coding agent, along with its foundation model Muse Spark 1.2. Muse Code operates on macOS and Linux, handling planning, implementation, and verification of changes across large code repositories. It features an asynchronous background agent that runs in parallel with the main agent, continuing information gathering and verification while the main agent focuses on overall policy and implementation. The tool records operations in a local event log, allowing state restoration after crashes. Commands include '/plan' for breaking tasks into execution plans, '/grill' for verifying plan weaknesses, and '/goal' for continuing until a specified goal is achieved. Muse Spark 1.2 is retrained from Muse Spark 1.1, with increased computational resources and an expanded training environment. Meta used rejection sampling to select superior examples from multiple execution histories, and Muse Spark 1.1 generated training data for the new model. In benchmarks, Muse Code with Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, ranking second after Claude Opus 5 and Claude Code's 86.7%. On DeepSWE 1.1, it scored 59.3%, ranking third. In Meta's internal evaluation of 440 development tasks, it scored 70.6%, exceeding the old model's 68.3%.
Muse Code is available from the terminal on macOS or Linux, and it goes beyond generating code fragments by investigating the entire repository, planning changes, and executing tests and verification. The asynchronous background agent runs in parallel, reporting results as needed, which allows the main agent to concentrate on high-level decisions. The local event log records instructions, tool executions, file edits, and user approvals, enabling recovery from crashes.
Muse Spark 1.2 is integrated into Muse Code and also available via the Meta Model API. Meta trained the model using rejection sampling, selecting superior examples from multiple execution histories, and used Muse Spark 1.1 to generate difficult coding tasks and instructions for training data. In the kernel acceleration experiments for NVIDIA Hopper GPUs, the combination of Muse Code and Muse Spark 1.2 executed over 1000 tool calls over up to 24 hours, achieving a 68.7% speedup for KDA and 61.1% for MLA, though it did not surpass Claude Opus 5's scores. Meta reduced duplicate computations by sharing compressed data for keys and values, and designed a Triton pipeline using kernel fusion and tiling.
Meta cautions that benchmark results do not strictly compare model performance alone, as the coding agents combined with each model differ. Muse Code is in beta, and Meta plans to expand agent environment features and deploy larger models in the future.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.