The model was trained mainly on long-running tasks like generating entire repositories or conducting independent research. To stay on track during these extended sessions, it plans steps ahead, works toward a fixed goal, and compresses its prior context instead of cutting it off.

Some of the training data came from the predecessor model itself. Muse Spark 1.1 generated programming tasks and instruction templates, then scored how well attempted solutions met the requirements. Meta says this process helps 1.2 follow complex instructions more accurately than its predecessor.

Meta cites Terminal-Bench 2.1, DeepSWE v1.1, and 440 tasks from its own codebase as evidence, comparing Spark 1.2 against Grok 4.5, Claude Opus 5, GPT-5.6 Terra (not OpenAI's stronger Sol model), and Gemini 3.6 Flash. Spark 1.2 shows a clear step up from Spark 1.1 but doesn't always close the gap to the top performers.

Meta's own methodology acknowledges that the test setup wasn't tuned for competing models and may not reflect their best results. A look at other leaderboards confirms this: Opus 5, for instance, scores about two percentage points higher elsewhere than in Meta's charts. Kimi K3 also shows up in Meta's methodology document. The company says it tested against the open model, but K3 is missing from the published benchmarks. On Terminal-Bench 2.1, K3 trails Opus 5 by only a slim margin and sits well ahead of Spark 1.2.

The DeepSWE runs also can't be compared directly to the official leaderboard because each model ran inside its own agent. That's because Meta now ships its own alternative to Claude Code and OpenAI's Codex alongside the new model.

The agent picks up exactly where it left off after a crash

Muse Code runs in the terminal and installs with a single command, just like Claude Code and OpenAI's Codex. The planning mode will feel familiar too: "/plan" generates a plan that the user has to approve. Less common is the counterpart "/grill," which stress-tests the plan for weaknesses before execution. There's also "/goal," which drives the agent toward a fixed objective.

The real new feature is how the tool handles sub-agents. Meta's helper agents stay active for an entire session and report back on their own when they have something to share, rather than spinning up for a subtask and shutting down afterward. This should cut down on redundant research.

Muse Code also has a fast-resume feature for crashes. The agent logs every model call, every approval, and every change to a local protocol file. Competitors keep full logs too, but Meta claims more precision: after a crash, Muse Code picks up exactly where it stopped instead of re-reading the full context.

Discount in exchange for training data

Wang says Meta is competing on price, not capabilities. The standard pricing matches version 1.1: a relatively cheap $1.25 per million input tokens and $4.25 per million output tokens. New is a tier at just 20 cents per million output tokens that requires users to share their data for model improvement.

Western competitors charge between $10 and $30 per million output tokens, while Chinese providers start at 18 cents. Kimi K3, the most conspicuous gap in Meta's benchmarks, costs $3 per million input tokens and $15 per million output tokens, though cache hits bring input down to just 30 cents.

Meta's stock dropped ten percent last week. The company generates 98 percent of its revenue from advertising.