Published: 6 August 2026
Meta Superintelligence Labs released Muse Code, a terminal coding agent in beta, alongside Muse Spark 1.2, the coding-focused model that powers it, on 5 August 2026. In Meta’s own evaluations Muse Spark 1.2 reaches 82.9 percent on Terminal-Bench 2.1 and 59.3 percent on DeepSWE v1.1, up from 76.2 and 53.0 percent for Muse Spark 1.1. Both are available now through the Meta Model API, and Muse Code installs on macOS and Linux with a single command.
What can Muse Code do?
Muse Code is a terminal coding agent built for software engineering work across large repositories. It plans changes, writes code and validates the results, and it can coordinate several persistent subagents on a single task. Unlike Anthropic’s Claude Code and OpenAI’s Codex, Muse Code ships without a desktop application and runs entirely from the terminal.
The runtime is built around a local append-only event log. Every model call, tool run, approval and edit is written to that log, which Meta describes as replay-exact and restart-safe: after a crash the agent resumes at the exact point where it stopped. That design is what allows Muse Code to run tasks lasting many hours without losing progress. Meta also ships a set of async background agents that stay alive for a whole session instead of being spawned per task, which avoids repeated information gathering and reduces the need for manual steering.
Three skills ship with the agent by default. The /plan command turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward completion of a stated objective. Installation is one line on macOS and Linux: curl -fsSL https://dev.meta.ai/install.sh | bash. There is no Windows build at launch, and the agent is distributed as a closed-source native binary named muse.
Muse Spark 1.2 benchmarks and technical specs
Muse Spark 1.2 has a context window of 1,048,576 tokens and accepts text, image, video and PDF input, returning text output. Meta publishes no parameter count and no architecture details. The API model IDs are muse-spark-1.2 and muse-spark-1.2-contributor, with muse-spark-1.1 still listed as the earlier checkpoint.
On Meta’s launch charts Muse Spark 1.2 scores:
- 82.9 percent on Terminal-Bench 2.1 against 76.2 percent for Muse Spark 1.1,
- 59.3 percent on DeepSWE v1.1 against 53.0 percent,
- 70.6 percent on Meta Internal Coding Bench against 68.3 percent.
- Terminal-Bench 2.1 covers all 89 tasks in the official release.
- DeepSWE v1.1 covers 113 tasks across 91 repositories and five languages: TypeScript, Go, Python, JavaScript and Rust.
- Meta Internal Coding Bench consists of 440 tasks derived from real pull requests in Meta’s own codebase. All figures are pass@1 averaged over five attempts, run in isolated Daytona cloud sandboxes.
Meta attributes the gains to three training choices. It significantly scaled training compute on coding tasks while widening environment diversity. It co-trained the model with the Muse Code harness using rejection-sampled trajectories and recipe optimisations for goals, context compaction and subagents. And it ran a self-improvement loop in which Muse Spark 1.1 generated challenging coding environments and instruction templates, then graded candidate solutions to build the training set for 1.2.
How does Muse Spark 1.2 compare to Claude Opus 5 and GPT-5.6 Terra?
Meta compared Muse Spark 1.2 with Muse Spark 1.1, Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash and Kimi K3. Each model was run inside its own agent product: Muse Code for Muse Spark 1.2, Claude Code for Claude Opus 5, Codex for GPT-5.6 Terra, Antigravity for Gemini 3.6 Flash, Grok Build for Grok 4.5 and Kimi Code for Kimi K3, each at its maximum reasoning setting.
On those charts Muse Spark 1.2 places second behind Claude Opus 5 on all three coding benchmarks: Terminal-Bench 2.1, DeepSWE v1.1 and Meta Internal Coding Bench. Meta’s methodology document states plainly that its evaluation setup, including agent tools and system prompts, may not be tuned to the strengths of third-party models, so the competitor figures may not reflect their best performance.
Every number in the launch is vendor-run, and no independent reproduction existed at the time of publication. Because the benchmark gains were measured inside the harness the model was co-trained with, results in a generic scaffold are likely to be lower.
Muse Spark 1.2 pricing and availability
Standard-tier pricing is unchanged from Muse Spark 1.1: $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Web search grounding costs $2.50 per 1,000 queries. Meta states there is no long-context premium and that injected steering tokens are not billed.
The second tier, muse-spark-1.2-contributor, costs $0.10 per million input tokens, $0.002 per million cached input tokens and $0.20 per million output tokens. That is a 12.5x discount on input and 21.25x on output, and the trade is explicit: Meta may use prompts and completions sent to that tier to train future models. The contributor tier is capped at 60 requests per minute per team against 3,000 on the standard tier, so it suits low-concurrency workloads only.
Muse Spark 1.2 is available through the Meta Model API, which Meta now describes as having expanded global access, and inside Muse Code. The weights are closed. There is no Hugging Face repository for Muse Spark or Muse Code and no self-hosting path, so evaluation habits carried over from the Llama era do not apply to this family.
Case study: GPU kernel optimisation over 24 hours
Meta published a case study in which Muse Spark 1.2 iteratively optimised GPU kernels over more than 1,000 tool calls in runs lasting up to 24 hours. Working inside Muse Code, the model wrote, compiled, profiled and progressively improved KDA and MLA kernels for NVIDIA Hopper GPUs against a provided baseline implementation.
For KDA the baseline was the FLA Triton implementation, and the model was barred from importing third-party kernel libraries such as FLA directly. Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, combining standard fusion and tiling with KDA-specific optimisations. For MLA the reference was a PyTorch implementation at batch size 1, 64 heads, sequence length 8192 and latent dimension 512, where the model built a two-kernel Triton pipeline that reuses the shared KV latent as both K and V.
Meta presents this as a capability demonstration rather than a benchmark result.
The full announcement, the benchmark charts and the three-page evaluation methodology on Muse Spark 1.2 and Muse Code can be found at the official Meta Research blog, with model and pricing documentation at the Meta Devs site.
