Muse Code Co-Trains the Harness — Spark 1.2 Alone Is Incomplete
Meta shipped Muse Code beta, a terminal coding agent with persistent background subagents and a replay-exact event log, powered by Muse Spark 1.2. Lab coding benches lead the post; live Arena snapshots list muse-spark-1.2 Elo around launch week, while 1.2-specific Artificial Analysis Index figures were still landing. Same-day cyber-eval breach is a separate story.
Meta is no longer only shipping chat/multimodal Muse — it is competing in the terminal agent market where Anthropic (Claude Code), OpenAI (Codex / ChatGPT Work), and Cursor-class tools already live. Co-training model + harness is the real product thesis: Spark 1.2 without Muse Code is incomplete.
Meta shipped Muse Code (beta) — a terminal coding agent — powered by Muse Spark 1.2, the newest Muse Spark tier. Install path: curl -fsSL https://dev.meta.ai/install.sh | bash (macOS/Linux). Muse Spark 1.2 is available in Muse Code and Meta Model API with expanded global access. The stack targets long-horizon multi-file / whole-repo engineering.
What shipped
| Piece | Detail (Meta primary) |
|---|---|
| Muse Code | Terminal agent: plan, implement, validate multi-file changes across large repos |
| Background agents | Persistent specialized subagents across a session (not one-shot spawns) |
| Runtime | Local event log of model calls, tools, approvals, edits — replay-exact, restart-safe |
| Bundled skills | /plan (approval-gated plan), /grill (stress-test plan), /goal (objective completion) |
| Muse Spark 1.2 | Coding-focused update to 1.1: more coding train compute, environment diversity; co-trained with Muse Code harness |
| Long-horizon | Whole-repo generation, large projects, auto-research; planning + goal conditioning + context compaction |
| Case study | GPU kernel optimization over 1,000+ tool calls (up to 24 hours) on Hopper KDA/MLA kernels |
Meta frames 1.2 as an interim step “toward the frontier, with larger and much more capable models on the way.”
Claims vs checks
Lab-reported: Meta publishes Terminal-Bench 2.1, DeepSWE 1.1, and internal coding-bench charts on the launch post (see methodology report linked from the blog). Kernel case studies claim progressive speedups vs. baseline Triton/PyTorch references under tool-call budgets.
Independent (as available near launch):
- LMArena / Arena: public leaderboard snapshots list muse-spark-1.2 (xHigh) with Elo-style scores (e.g. ~1290±18 on a Hugging Face Arena snapshot around launch week) — treat as live board, not a frozen vendor claim.
- Artificial Analysis: original Muse Spark placed in the high tier of the Intelligence Index at base launch (mid-50s historically; base Muse Spark often cited ~52 near first GA). 1.2-specific full Index breakdowns were still landing around launch day — prefer Meta coding benches + any same-day AA posts over inventing a 1.2 Index number.
- Prior AA coding-index snapshots had base Muse Spark lagging GPT-5.4 / Gemini 3.1 Pro / top Claude coding configs — 1.2 is Meta’s explicit answer to that gap.
Distinct from Muse Image, Muse Spark 1.1 (July 9), and Llama open weights. Same-day Meta cyber-eval breach disclosure (Muse Spark via Irregular) is a separate safety story.
Limits
- If Arena votes for 1.2 are thin in the first days, rank is not settled.
- 1.2-specific full AA Index may lag the launch post.
- API pricing/rate limits and “larger models on the way” remain open.
Sources
- Meta AI Research: Introducing Muse Code and Muse Spark 1.2 (August 5, 2026)
- Meta AI Research hub
- LMArena / Hugging Face arena leaderboard (live; check muse-spark-1.2)
- Artificial Analysis: Muse Spark model page
- Times of AI:
meta-muse-spark-1-1-model-api(July 9)