Meta has released Muse Code, a terminal-based AI coding agent now in beta, alongside Muse Spark 1.2, a coding-focused update to its Muse Spark model family. The dual launch positions the company against Anthropic’s Claude Code, OpenAI’s Codex, and other agentic coding tools that have become central to how many professional developers ship software.
Mark Zuckerberg announced the release on X: “Releasing Muse Code in beta today. It’s a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results.”
Until now, Meta’s developer efforts centered on the open-weight Llama models. Muse Code marks a shift. It is a full proprietary harness, installable on macOS or Linux with a single command (`curl -fsSL https://dev.meta.ai/install.sh | bash`), and is co-trained with the model that powers it. Users must log in with a Meta account and provide billing details before it will run.
Persistent background agents and parallel worktrees
Muse Code’s core architectural choice is persistent async background agents. Instead of spawning fresh helpers for every task, specialized agents stay alive for the entire session. Meta says this reduces redundant repository exploration and lowers latency. For larger jobs, the system fans out to parallel sub-agents, each running in its own isolated git worktree so the developer’s working copy remains untouched. Zuckerberg noted that in testing it built six features for a game simultaneously without collisions.
Every model call, tool run, approval, and edit is written to a local event log before execution. Meta describes the runtime as “replay-exact and restart-safe.” If the agent crashes deep into a long-running task, it can resume exactly where it left off.
The product also ships with familiar skills: `/plan` produces an approval-gated plan, `/grill` stress-tests it, and `/goal` drives the agent toward a stated objective.
Muse Spark 1.2: co-trained with its harness
Muse Spark 1.2 is described as a coding-focused update to Muse Spark 1.1 with significantly more training compute on coding tasks and broader environment diversity. Gains are claimed in code generation, complex debugging, and codebase understanding while retaining general agentic ability.
Two training details stand out. First, the model was co-trained with Muse Code itself using rejection-sampled harness trajectories and optimizations for goals, context compaction, and sub-agents. Second, Meta ran a self-improvement loop in which Muse Spark 1.1 generated challenging coding environments and graded candidate solutions, producing scalable training data for 1.2.
Benchmarks
Meta published comparisons on Terminal-Bench 2.1, DeepSWE 1.1, and an internal coding benchmark. Results show solid second- or third-place finishes rather than leadership:
- Terminal-Bench 2.1: Muse Spark 1.2 + Muse Code scored 82.9%, ahead of GPT-5.6 Terra in Codex (81.8%) and Grok 4.5 in Grok Build (81.6%), but behind Anthropic’s Opus 5 at max effort in Claude Code (86.7%).
- DeepSWE 1.1: 59.3% (third behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%).
- Internal Meta benchmark: 70.6%, ahead of GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%), but nearly nine points behind Opus 5 (79.4%).
Claude models topped all three charts. Muse Spark 1.2 improved on 1.1 by 6.7 points on Terminal-Bench and 6.3 on DeepSWE; some of that gain is attributable to the new harness rather than the model alone, since 1.1 scores were recorded in a generic mini-swe-agent environment.
Meta also highlighted a long-horizon demonstration: Muse Spark 1.2 spent more than 1,000 tool calls over up to 24 hours optimizing GPU kernels on NVIDIA Hopper hardware. Working in Triton and barred from wrapping existing libraries, it produced substantial improvements over baseline KDA and MLA kernels, including non-obvious optimizations such as re-centering gated cumulative decay at a chunk midpoint. Zuckerberg said the agent continued finding meaningful gains well beyond the initial exploration phase.
Pricing: standard vs. contributor tiers
Muse Spark 1.2 is available through the Meta Model API in two tiers.
- **Standard**: $1.25 per million input tokens / $4.25 per million output tokens (cached input $0.15). Prompts and completions are not used for training. Rate limits reach 3,000 requests and 4 million tokens per minute per team. There is no long-context premium.
- **Contributor**: $0.10 / $0.20 (cached input $0.002)—roughly 12–21× cheaper—in exchange for explicit permission to use prompts and completions for training future models. Rate limits are much tighter (60 requests per minute). Zuckerberg has steered new users toward this tier, calling it “easy and low-cost to get started.”
Even the contributor tier requires a payment method on file; the agent will not run until billing is set up. Meta frames the low-cost option as suitable for prototyping where training on user data is acceptable. Enterprises with proprietary codebases will need to opt into the standard tier to keep their data out of Meta’s training pipeline.
The open-source question
The announcement contains no mention of open weights or self-hosting—a notable departure from Meta’s earlier Llama strategy. After years of positioning open-source AI as “the path forward,” Meta shipped the original Muse Spark in April as its first proprietary, cloud-only model. Asked whether Muse Code would be open-sourced, Zuckerberg replied only that he would “have more to share on that soon.”
Rivals have moved in the opposite direction: OpenAI open-sourced its Codex CLI under Apache 2.0, and Google’s Gemini CLI is likewise Apache-licensed. Muse Code’s posture is closer to Anthropic’s proprietary Claude Code. The contributor tier effectively replaces the old Llama flywheel (free weights for mindshare) with cheap tokens in exchange for training data.
Why it matters
Terminal coding agents have become one of the fastest-growing surfaces in enterprise AI. Meta’s entry brings a distinct architecture—persistent background agents plus an append-only local event log—a credible long-horizon demo, and an aggressive pricing wedge. The open questions remain the ones benchmarks cannot settle: whether Muse Spark 1.2 matches Claude- and GPT-class performance on real-world repositories, whether developers will trust Meta with their code, and whether the contributor discount is large enough to overcome those concerns.
Muse Code is available in beta today. Muse Spark 1.2 is live in the Meta Model API with expanded global access.
