Meta launches Muse Code beta alongside Muse Spark 1.2

Meta just put a coding agent in the terminal and priced the model behind it like a product, not a research teaser. Per Meta AI Research, Muse Code is in beta for macOS and Linux on Muse Spark 1.2, with the same checkpoint on the Meta Model API under expanded global access. One curl install. Token-metered bills.

What Muse Code can run in the terminal

Muse Code is not a chat pane with a generate-function button. Meta describes a local agent that plans changes across large repositories, writes code, and validates results. It coordinates multiple persistent subagents per task instead of spawning throwaway helpers that re-gather context every turn.

The runtime is concrete. Every model call, tool run, approval, and edit lands in a local event log. That log is the source of truth, so a crash can resume where work stopped. Long-running jobs stop dying with the shell session. Bundled skills include /plan for approval-gated plans, /grill to stress-test those plans, and /goal to drive toward a stated objective. Async background agents stay alive for the session, pick next steps, and decide when to ping the main loop.

What Spark 1.2 changes vs 1.1

Spark 1.2 is a coding-focused update to 1.1: more training compute on coding tasks, wider training environments, better debugging and codebase understanding, while Meta says general-agent strength holds. The interesting bit is co-training. Meta trained the model with Muse Code in the loop, using rejection-sampled harness trajectories and recipe work for goals, compaction, and subagents so the weights and the toolset fit each other.

Long-horizon work is the pitch: whole-repository generation, large end-to-end projects, and auto-research loops that plan, stay goal-conditioned, and compact context. Meta also used Spark 1.1 to invent hard coding environments, then graded candidates to feed 1.2. Vendor framing until independent evals land.

A kernel case study is the showpiece. On NVIDIA Hopper, Muse Code plus Spark 1.2 spent 1,000-plus tool calls over as much as 24 hours writing, compiling, and profiling Triton kernels for KDA and MLA against a baseline, without importing third-party kernel libraries. Useful if it holds for your stack. Not a substitute for your own bench.

API access and pricing signals

Spark 1.2 ships in Muse Code and on the Meta Model API. Meta’s pricing docs list Standard at $1.25 / $0.15 cached / $4.25 output per million tokens, with prompts and completions not used for training. Contributor (muse-spark-1.2-contributor) drops that to $0.10 / $0.002 / $0.20, in exchange for training rights on your traffic. Rate limits diverge: Standard at 3,000 RPM and 4M TPM; Contributor at 100 RPM and 3M TPM. The cheap lane is a data deal with a throttle.

Who should beta-test first

Start here if you already live in a terminal agent and can isolate a non-sensitive repo for burn-in. Replay-safe long jobs matter most for multi-hour refactors, migrations, and systems work where restart cost is the real tax. Keep NDA code on Standard, or off Meta entirely, until legal signs the Contributor bargain. High-concurrency teams should not pretend Contributor’s discount survives a 100 RPM ceiling.

Analysis: Meta is late to the terminal-agent party and early to admitting the agent and the model have to be co-trained as a product. If Spark 1.2’s long-horizon demos hold outside Meta’s kernels, Muse Code becomes a serious third lane. If the beta stalls on ordinary app repos, this is another research blog with a curl installer. Watch independent SWE-bench-class runs and whether Contributor becomes Meta’s quiet fine-tuning pipeline.