Muse Spark 1.3’s 75.4 DeepSWE score is a strong Meta-run result

Meta released Muse Spark 1.3 on September 2, 2026. Its comparison card lists 75.4% on DeepSWE v1.1. As of September 4, that score is still missing from DeepSWE’s public leaderboard. Both can be true: the model may have improved on long-horizon software engineering, while 75.4% remains a vendor-run evaluation rather than an independently listed entry.

That distinction matters more than the headline. Version, reasoning effort, agent harness, and who ran the eval all change what the number actually means.

The release is bigger than one row

Muse Spark 1.3 is rolling out in Muse Code and the Meta Model API. Current reasoning modes are live; Meta says max will follow more safety testing. That timing is important because the launch card labels the DeepSWE run as max. The top chart configuration may not yet match what teams can use in production.

The product pitch is long-running, multi-tool work: recover missing context, revise a plan when sources conflict, ask for help when blocked, and confirm before high-impact actions. Internally, Meta also reported about 20% fewer tool calls and 25% fewer tokens versus 1.2, with fewer idle turns and less verbose code. Those are useful product claims, not guaranteed savings on every repo, toolset, or review policy.

75.4% is real. Provenance still belongs next to it.

Meta generated Muse Spark 1.3 scores through its own API. Competitor numbers mix Meta runs, official boards, and provider self-reports. Meta picked the highest comparable primary metric and noted that third-party settings were best-effort, not necessarily provider-optimized.

DeepSWE v1.1 is 113 long-horizon tasks across 91 repositories and five languages. A task passes only if handwritten functional and regression tests both pass. Meta ran 1.3 at max with mini-swe-agent and pulled the other scores from Datacurve’s official board. On September 4 that board showed Gemini 3.8 Flash high and Claude Opus 5 max at 74%, then GPT-5.6 Sol max at 73%. Muse Spark 1.3 was not listed. The board also publishes uncertainty ranges from repeated runs; Meta’s 75.4% does not.

The accurate sentence is: Meta’s run scored above the current public leaders. It is not accurate to say DeepSWE independently verified Muse Spark 1.3 as first.

One scorecard, three kinds of comparison

The full table is still useful. It shows a large jump over 1.2 and maps strengths. Each row answers a different question:

SWE-Atlas is a reminder that harness choice belongs in every comparison: Scale lists mini-swe-agent next to native scaffolds such as Claude Code and Codex CLI, and notes that native scaffolds can raise scores. Terminal-Bench makes the same point by splitting agent, model, and effort into separate columns.

The rest of the card is mixed, not a sweep. Opus 5 remains higher on GDPval-AA v2, JobBench, OSWorld 2.0, and AutomationBench. GPT-5.6 Sol leads DeepSearchQA and Meta’s internal Agentic IF Index. Terminal-Bench 2.1 is a tie at 88.8%, not an outright win. AutomationBench is 49.4%.

Lock four variables before you migrate

Record these next to every result:

1. Model version — pin muse-spark-1.3. A product label is not a checkpoint.
2. Reasoning effort — test each live setting. Do not treat future max as current production.
3. Agent harness — Muse Code, mini-swe-agent, or your internal tools, including permissions, timeouts, retries, and available tools.
4. Provenance — vendor-run, public-board, third-party reproduction, or internal regression. Label them separately.

Then run 30–100 tasks from real work. Hold container, prompt, tools, and acceptance tests fixed. Track pass rate, tool calls, tokens, P50/P95 latency, retries, human rework, and:

cost per accepted task = total model and tool spend / tasks that pass acceptance tests

Migrate if the 20% tool-call and 25% token claims survive that test. A one-point public-benchmark edge is worthless if latency, refusals, or review time go up.

Who should test first

Teams already on Muse Code or the Meta Model API have the cleanest path, especially if they do repo-wide changes, long plans, multi-tool work, and tight constraints. Version 1.2 is the right baseline.

Short completions, summaries, and single-file edits do not need a migration because of 75.4%. Teams waiting for a public listing, general max access, or downloadable weights can stay put. Meta has an open-weights Muse Spark release on the roadmap, without a date and without confirming it will be this 1.3 checkpoint.

FAQ, compressed

Available now? Yes, in Muse Code and the Meta Model API. Current modes are live; max comes after more safety testing.

First on DeepSWE? Meta reports 75.4%, above the 74% public leaders on September 4. The board does not list 1.3. Call it a Meta-run result above current leaders, not a verified public win.

Comparable to the public board? Both used mini-swe-agent, so it is closer than a cross-harness test. Meta still ran 1.3 itself, pulled the rest from the board, and omitted an uncertainty range. Wait for a listing or reproduce before treating a narrow lead as decisive.

Downloadable weights? Not now. Official access is Muse Code and the Meta Model API.

The real takeaway

The 1.4-point gap between 75.4 and 74 is not the story. Vendors now ship a coupled model, reasoning mode, and agent system. That is closer to real development — and it makes “best model” an incomplete phrase. Teams need the best reproducible system for their repo, controls, and budget.