Background
<cite index="1-1">Muse Spark is a Large Language Model (LLM) developed by Meta through its Meta Superintelligence Labs (MSL).</cite> <cite index="6-3">The first model in the series was announced on April 8, 2026, as the first in a new series of LLMs built by MSL.</cite> <cite index="10-2">Muse Spark was described at launch as a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration.</cite> <cite index="2-3,2-4">It is Meta's first proprietary, closed AI model, and its release marked the end of Meta's open-weight Llama era.</cite>
The Muse Spark 1.2 Release
<cite index="26-1">Muse Spark 1.2, released August 5, 2026, is Meta Superintelligence Labs' coding-focused update to the July 9 Muse Spark 1.1, shipped alongside Muse Code, a beta terminal coding agent that is co-trained with the model.</cite> <cite index="22-2,22-3">The company's official research blog framed Muse Code (beta) as Meta's "next step toward the frontier, with larger and much more capable models on the way."</cite>
<cite index="2-11,2-12">Meta has described Muse Spark 1.2 as a coding-focused point release of 1.1, not a new base model, and states it was co-trained with the new coding agent using rejection-sampled harness trajectories, plus a self-improvement loop in which Muse Spark 1.1 generated training environments and graded candidate solutions for 1.2.</cite>
Muse Code
<cite index="22-4,22-5">According to Meta's research blog, Muse Code takes on complex software engineering tasks across large repositories — planning changes, writing code, and validating results — and can coordinate multiple persistent subagents for each task.</cite> <cite index="29-13">The terminal tool is available for macOS and Linux and is built around persistent background agents and parallel sub-agents in isolated Git worktrees.</cite>
Specifications and Availability
<cite index="5-2,5-3,5-4">Muse Spark 1.2 is multimodal, capable of processing text and image input and generating text output, with a context window of 1.0 million tokens.</cite> <cite index="5-6,5-7,5-8">The model is proprietary; the weights are not publicly available, and Meta has not disclosed the model size or parameter count.</cite>
<cite index="26-3,26-4">Standard-tier API pricing is unchanged from Muse Spark 1.1 at $1.25 input and $4.25 output per one million tokens, while a new "contributor" tier offers $0.10 input and $0.20 output per million tokens — a discount of approximately 12.5× on input — in exchange for permission to use submitted data to train future Meta models.</cite> <cite index="1-14">Muse Spark 1.2 was made available through the Meta Model API and OpenRouter.</cite>
Benchmarks and Competitive Position
<cite index="23-1">Meta reported that Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1.</cite> <cite index="25-1">On Meta's own benchmark charts, Muse Spark 1.2 trails Claude Opus 5, which Meta places at 86.7% on Terminal-Bench 2.1.</cite> Independent evaluators have noted methodological caveats: <cite index="23-11,23-12">both figures come from Meta's own evaluation harness, and neither Muse Spark 1.2 nor Claude Opus 5 held an independently verified entry on the official Terminal-Bench leaderboard as of the release date.</cite>
<cite index="21-3,21-4">Independent benchmarker Artificial Analysis scored Muse Spark 1.2 at 54 on its Intelligence Index — Meta's third release in four months — up 3 points from Muse Spark 1.1 (51) and 11 points from the original Muse Spark 1.0 (43) in April.</cite> <cite index="21-5">That score places it effectively tied with GPT-5.5 (55) and Grok 4.5 (54), narrowly behind current frontier leaders including Claude Opus 5 (61) and GPT-5.6 Sol (59).</cite>
Broader Context
The August 5 release continues a pattern of accelerating model cadence across major laboratories in 2026. <cite index="28-4,28-5">Meta released Muse Spark 1.2 four weeks after Muse Spark 1.1 and roughly four months after the first Muse Spark, with Muse Code marking the first time Meta shipped a dedicated coding agent alongside a model update.</cite> The aggressive pricing on the contributor tier is being watched as a potential market-share lever, with independent analysis noting <cite index="27-6">the contributor pricing tier at $0.10/$0.20 per million tokens represents the most aggressive pricing from any capable coding model available at the time of release.</cite>