Meta's Muse Code + Spark 1.2: A Terminal Agent Built for Whole Repositories

On 5 August 2026, Meta's Superintelligence Labs shipped two things at once: Muse Code, a beta terminal coding agent, and Muse Spark 1.2, the model built to run it. That pairing is the whole point. This isn't "here's a bigger model, go wire up your own harness" — it's a model and an agent co-trained together so they behave as one system. For anyone who's watched Claude Code and Copilot's whole-feature push define the terminal-agent category in 2026, this is Meta planting a serious flag.

1Mtoken context window

82.9%Terminal-Bench 2.1

59.3%DeepSWE v1.1

2parts, co-trained as one

What actually shipped

Muse Spark 1.2 is the model: a 1M-token context window meant to hold dependency graphs, legacy code, and thousands of files in a single session. Meta's own numbers put it at 82.9% on Terminal-Bench 2.1 (up from 76.2% for Spark 1.1) and 59.3% on DeepSWE v1.1 (up from 53.0%). Those are self-reported, so weight them accordingly — but the deltas are the interesting part: this is an iterative, harness-aware bump, not a from-scratch flagship.

Muse Code is the agent: a terminal tool in the same lineage as Claude Code and Codex, with a 12× cheaper "contributor" tier aimed at getting it into a lot of hands fast.

The design thesis is in three model-level features that matter more than the benchmark line:

Context compaction — the model actively compresses what it's learned so a long task doesn't drown in its own history. This is the bigger-context-window-is-a-design-problem insight made into a feature: a 1M window is useless if the model can't decide what to keep.
Goal conditioning — mechanisms to keep the agent pointed at the original objective across dozens of tool calls, instead of drifting into a plausible-but-wrong side quest.
Planning + subagents — sequencing long work and spawning focused sub-tasks, trained end-to-end rather than bolted on by a prompt.

The real story: whole-repository training

Here's what separates this from a normal model release. Meta says Spark 1.2 was trained on whole-repository generation, large end-to-end projects, and auto-research — with rejection-sampled harness trajectories and recipe tuning specifically for goals, compaction, and subagents. Then it was co-trained with Muse Code so the model's best behavior shows up precisely when paired with that agent.

That's a meaningful shift in how these things get built:

Old approach	Muse's approach
Train a general model	Train on whole-repo, long-horizon tasks
Bolt an agent on top with prompts	Co-train model + agent as one system
Big context, hope it's used well	Context compaction as a trained skill
Benchmark on isolated problems	Optimize the harness trajectory itself

The bet is that the agent harness is part of the model, not a wrapper around it. If that's right, "which model" and "which agent" stop being separable questions — and the moat moves from raw weights to the co-trained model-plus-harness system. That's the same lesson the editor-as-training-pipeline framing pointed at, taken one step further.

Should you use it?

Honest read for a working engineer:

If you live in Claude Code or Codex today, Muse Code is worth a spin specifically for large, messy repos — the whole-repo training and compaction are aimed exactly at the "my codebase doesn't fit in the agent's head" problem. The 12× cheaper tier lowers the cost of trying.
If you need it in production, wait. It's beta, the benchmarks are first-party, and Terminal-Bench/DeepSWE numbers don't capture how a tool feels on your code. Treat the scores as an invitation to test, not a verdict.
If you're picking a stack, note the strategic point: Meta is now a third serious terminal-agent player alongside Anthropic and GitHub. That's good for you — competition here means faster iteration and cheaper tiers. It's also a reason to keep your workflow portable rather than betting the farm on one agent's ergonomics.

The honest caveats

All benchmarks are Meta's own. No independent Terminal-Bench 2.1 verification yet. Self-reported gains are directionally useful, not gospel.
Beta means beta. Expect rough edges, changing pricing, and API churn. The 12× cheaper contributor tier is a land-grab price, not a guaranteed floor.
"Co-trained" is a claim, not a measurement. The idea is compelling; whether it actually beats a good general model + a good generic harness is something the field will sort out over the next few months.

My take

The number that matters isn't 82.9% — it's the word co-trained. For two years we treated the model and the agent as separate layers: pick a smart model, wrap it in a clever harness. Meta is betting that's the wrong seam, that the harness trajectory is training data and the agent is part of the model. If they're right, the category consolidates around integrated systems and the "best model" question gets a lot less meaningful on its own. If they're wrong, it's a solid iterative release with good marketing. Either way, three credible terminal agents competing on price and capability is exactly the market a working engineer wants.

Sources

Benchmarks are Meta's own as reported on 5 August 2026; independent evaluation pending. Spot a corrected detail? Tell me and I'll update.