Meta Muse Code & Muse Spark 1.2: Terminal Coding Agent, Benchmarks & What It Means for AI Developers (2026)

August 5, 2026 — Meta AI Research released Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, the company’s newest coding-focused model. Muse Code plans changes across large repositories, writes code, validates results, and coordinates persistent background subagents — positioning Meta as a direct competitor in the agentic coding space alongside tools like Cursor, Claude Code, and GitHub Copilot Workspace.

This guide breaks down what Muse Code and Muse Spark 1.2 are, how they work under the hood, benchmark claims, how to install Muse Code on macOS or Linux, and what the release signals for Meta’s broader AI developer strategy. Source: official Meta AI Research announcement.

Too long? Start here

• Muse Code (beta): Meta’s terminal coding agent — install via curl -fsSL https://dev.meta.ai/install.sh | bash
• Muse Spark 1.2: Coding-focused model update with stronger generation, debugging, and repo understanding
• Key architecture: Main agent loop + persistent async background subagents that stay active across a session
• Crash recovery: Local event log makes sessions replay-exact and restart-safe
• Built-in skills: /plan (approval-gated planning), /grill (stress-test plans), /goal (objective completion)
• Availability: Muse Spark 1.2 in Muse Code and Meta Model API with expanded global access

What is Muse Code?

Muse Code is Meta’s answer to the rise of agentic coding assistants — tools that do more than autocomplete a function. It runs in the terminal and takes on complex software engineering work: planning multi-file changes, implementing features, debugging failures, and validating outcomes across large codebases.

Unlike one-shot chat completions, Muse Code is designed for long-horizon tasks that may span hundreds or thousands of tool calls. Meta demonstrated this with a kernel optimization case study where the agent ran 1,000+ tool calls over up to 24 hours, iteratively writing, compiling, and profiling GPU code.

Meta also showed a creative demo: a user drops an MP4 fly-through video of a vacation home into the terminal, and Muse Code interprets the footage to produce a marketing and booking page — evidence the agent loop extends beyond pure backend engineering.

How Muse Code works: agents, runtime, and skills

Async background agents

Muse Code uses a simple main agent loop augmented by persistent async background agents. These specialized subagents remain active throughout the entire session rather than being spawned and torn down per task.

That persistence matters for real engineering work:

• Background agents avoid redundant codebase exploration on every subtask
• They carry out next steps independently and decide when to report back to the main agent
• Latency drops on difficult, multi-step problems because context is retained
• Users need less manual steering mid-task

This design mirrors trends across the industry — specialized subagents for exploration, testing, and review — but Meta emphasizes session-long persistence as a differentiator.

Replay-exact runtime design

Every model call, tool run, approval, and edit is appended to a local event log. That log is the single source of truth for the session.

Consequences for developers:

• Replay-exact: Sessions can be reconstructed step by step
• Restart-safe: After a crash, Muse Code resumes precisely where it stopped
• Long-running tasks: Multi-hour or multi-day jobs are viable without losing progress to transient failures

For teams running expensive agent workloads — kernel tuning, large refactors, research loops — crash recovery is not a nice-to-have; it is infrastructure.

Bundled skills: /plan, /grill, /goal

Muse Code ships with default skills that shape how work gets done:

The approval gate on /plan is particularly notable: it inserts a human checkpoint before the agent commits to a direction — a pattern enterprise teams often require for production codebases.

Muse Spark 1.2: the model behind Muse Code

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. Meta scaled training compute on coding tasks and expanded training environment diversity while keeping strength in general agent capabilities.

Improvement areas

• Code generation — cleaner, more context-aware output across languages and frameworks
• Complex debugging — tracing failures across files and build systems
• Codebase understanding — navigating large monorepos and unfamiliar structures
• End-to-end developer workflows — from plan to implement to validate without losing thread

Co-training with Muse Code

Meta co-trained Muse Spark 1.2 alongside the Muse Code harness. Training included:

• Rejection-sampled harness trajectories (keeping high-quality agent runs)
• Recipe optimizations for goals, context compaction, and subagent coordination
• Integration of the Muse Code toolset for maximum harness compatibility

The result: the model is tuned not just for isolated coding prompts, but for how it behaves inside Muse Code’s agent loop.

Long-horizon training

Muse Spark 1.2 was trained extensively on long-horizon tasks:

• Whole-repository generation
• Large end-to-end projects
• Auto-research workflows

It uses planning to sequence work, goal conditioning to stay on objective, and context compaction to retain critical knowledge as sessions grow — the same primitives Meta needs for agents that run for hours, not seconds.

Self-improvement loop

Meta used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates, then had the model grade candidate solutions against those requirements. That scalable synthetic dataset helped Muse Spark 1.2 follow complex instructions more precisely than its predecessor — a classic RLHF-style flywheel applied to agentic coding.

Benchmarks and the kernel optimization case study

Meta published benchmark comparisons on Terminal-Bench 2.1, DeepSWE 1.1, and an internal Meta Coding Bench — positioning Muse Spark 1.2 competitively against other leading coding models. Full evaluation details are in Meta’s technical report linked from the announcement.

GPU kernel optimization (1,000+ tool calls)

The most striking demo is iterative GPU kernel optimization on NVIDIA Hopper hardware:

• Agent writes Triton kernels, compiles, profiles, and improves over the baseline
• Runs up to 1,000+ tool calls across sessions lasting up to 24 hours
• Benchmarked on KDA and MLA kernel workloads

Models were prohibited from importing third-party kernel libraries directly — they had to implement algorithms in Triton using specialized optimization knowledge, not wrap existing implementations.

KDA highlights: Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, using KDA-specific optimizations like re-centering the gated cumulative decay at the chunk midpoint.

MLA highlights: A two-kernel Triton pipeline reusing the shared KV latent as both K and V — benchmarked at batch size 1, 64 heads, sequence length 8192, latent dimension 512.

These results matter because they show Muse Spark 1.2 can sustain iterative scientific engineering, not just boilerplate CRUD — the bar for “coding agent” is moving from autocomplete to autonomous optimization loops.

How to install Muse Code

Meta provides a one-line installer for macOS and Linux:

curl -fsSL https://dev.meta.ai/install.sh | bash

After installation, Muse Code runs in the terminal against your local repositories. Muse Spark 1.2 is the default model powering the agent.

Availability and Meta Model API

As of August 2026:

• Muse Code (beta): Available via the install script above
• Muse Spark 1.2: Available inside Muse Code and through the Meta Model API with expanded global access

Meta signals more harness features and larger models are on the horizon — this release is positioned as a step toward the frontier, not the final form.

Muse Code vs other coding agents

Muse Code is not replacing IDE workflows — it targets engineers who want a terminal-native, long-running agent with Meta’s model stack and crash-safe runtime.

What this means for Meta’s AI strategy

Muse Code and Muse Spark 1.2 sit alongside Meta’s consumer AI (Meta AI assistant), media generation (Muse Image), and business messaging AI (WhatsApp AI tools, Llama integrations). For developers, three signals stand out:

• Agent infrastructure is product: Meta is shipping harness + model together, not just weights
• Long-horizon is the benchmark: Training and evals emphasize multi-hour tasks, not single-file edits
• Global API expansion: Muse Spark 1.2 in Meta Model API lowers the barrier for teams building custom agents on Meta’s stack

For businesses already on Meta’s platform — WhatsApp Business API, Instagram, Facebook — stronger coding and agent tooling upstream often translates into better integrations, faster custom automation, and more capable in-house AI teams. See our guides on AI agents for business and WhatsApp AI agents for the customer-facing side of the same trend.

FAQ