# Muse Spark 1.3 Scores 75.4, but It Is Still Missing From DeepSWE's Public Leaderboard

**Meta released Muse Spark 1.3 on September 2, 2026, and its comparison card reports a 75.4% score on DeepSWE v1.1.** As of September 4, DeepSWE's public leaderboard does not include that result. Both facts can be true: Meta may have made a meaningful leap in long-horizon software engineering, while 75.4% remains a vendor-run evaluation rather than an independently listed leaderboard entry.

The distinction changes how engineering teams should read the launch. Model version, reasoning effort, agent harness, and result provenance matter more than the claim that one model ranks first.

> Editor's note: This article was rewritten on September 4, 2026, using Meta's release page, evaluation report, and the benchmark providers' public leaderboards. The first version linked to dead sources and described a Meta-run comparison as a third-party leaderboard result. Those errors have been corrected.

![Official Muse Spark 1.3 identity artwork.](https://s4.tenten.co/learning/content/images/2026/09/landing-page-1-5.png align="center")

#### This Release Changes More Than a Benchmark Row

Muse Spark 1.3 is rolling out through Muse Code and the Meta Model API. Existing reasoning modes are available now. Meta says `max` reasoning will arrive after additional safety testing.

That timing matters because Meta's comparison card labels Muse Spark 1.3 as a `max` run. The highest configuration in the launch chart may not yet match the mode a developer can use in production.

Meta focused the release on long-running, multi-tool work. The model is designed to recover missing context, revise a plan when sources conflict, ask users for help when blocked, and seek confirmation before consequential actions. Meta's internal comparison also found that version 1.3 used about 20% fewer tool calls and 25% fewer tokens than version 1.2, with fewer unnecessary turns and less verbose code.

Those efficiency figures are useful product claims, but they are still vendor measurements. They create a testable hypothesis. If the same workload consumes one-quarter fewer tokens without lowering acceptance rate, the new model may reduce cost per completed task. The percentages do not guarantee the same savings across every repository, toolset, and approval policy.

![Official Muse Spark 1.3 agent output showing an X-Wing flow-simulation report.](https://s4.tenten.co/learning/content/images/2026/09/landing-page-2-4.png align="center")

#### The 75.4% Result Is Real, and Its Provenance Belongs Beside It

Meta's methodology explains how it assembled the comparison. Muse Spark 1.3 results were generated through the Meta Model API. Competitor numbers may come from Meta's own runs, an official leaderboard, or a model provider's self-reported result. Meta selected the highest comparable primary metric and cautioned that its third-party model settings were best-effort rather than necessarily provider-optimized.

The DeepSWE row is more specific. DeepSWE v1.1 contains 113 long-horizon tasks across 91 repositories and five programming languages. A task only passes when its handwritten functional and regression tests both succeed. Meta ran Muse Spark 1.3 at `max` with mini-swe-agent and obtained the other model scores from Datacurve's official leaderboard.

On September 4, that public board listed Gemini 3.8 Flash `high` and Claude Opus 5 `max` at 74%, followed by GPT-5.6 Sol `max` at 73%. Muse Spark 1.3 was absent. The board states that every listed model uses mini-swe-agent for consistency and publishes uncertainty ranges from repeated runs. Meta's single 75.4% figure does not show a corresponding confidence interval.

The accurate claim is that Meta's run scored above the current public-board leaders. It is not yet accurate to say DeepSWE independently verified Muse Spark 1.3 as the winner. That is a reproducibility distinction, not a semantic technicality.

#### One Card Contains Three Different Kinds of Comparison

Meta's full scorecard remains valuable. It shows a large gain over Muse Spark 1.2 and provides a useful map of the model's strengths. Each row still answers a different question.

| Evaluation | Muse Spark 1.3 | What the score supports | What it does not yet support |
| --- | --- | --- | --- |
| DeepSWE v1.1 | 75.4% | A strong Meta-run result with mini-swe-agent at `max` | An independently listed DeepSWE leaderboard win |
| SWE-Atlas Codebase QnA | 59.4% | A gain over version 1.2 on repository comprehension | A harness-free comparison with native Claude Code or Codex runs |
| Terminal-Bench 2.1 | 88.8% | A tie with GPT-5.6 Sol in Meta's internal agent-evaluation framework | A submission already accepted by Terminal-Bench's verified board |
| MRCR 512K-1M | 98.1% | Strong target-string retrieval in that long-context setup | Reliable comprehension of any million-token repository |

SWE-Atlas demonstrates why the harness belongs in every model comparison. Scale's public page lists mini-swe-agent alongside native scaffolds such as Claude Code and Codex CLI, and it notes that native scaffolds can improve results. Terminal-Bench makes the dependency explicit by separating agent, model, and effort into different leaderboard columns.

![Official Meta comparison card for Muse Spark 1.3 evaluations.](https://s4.tenten.co/learning/content/images/2026/09/landing-page-3-3.png align="center")

The official card also corrects two important readings from the earlier article. Muse Spark 1.3 scores 88.8% on Terminal-Bench 2.1, tying GPT-5.6 Sol rather than winning outright. Its AutomationBench score is 49.4%, not 47.2%.

The broader table is a mixed profile, not evidence of a clean sweep. Opus 5 remains higher on GDPVal-AA v2, JobBench, OSWorld 2.0, and AutomationBench. GPT-5.6 Sol leads on DeepSearchQA and Meta's internal Agentic IF Index.

#### Lock Four Variables Before You Migrate

An engineering team can turn the launch into a useful internal evaluation by recording four fields beside every result.

1.  **Model version:** Pin `muse-spark-1.3`. Do not treat a continuously updated product label as a reproducible checkpoint.
    
2.  **Reasoning effort:** Test each available setting separately. Do not use a future `max` configuration as a proxy for current production behavior.
    
3.  **Agent harness:** Record Muse Code, mini-swe-agent, or the internal tool layer, including permissions, timeouts, retries, and available tools.
    
4.  **Result provenance:** Label vendor-run, public-board, third-party reproduction, and internal regression results separately.
    

Build a set of 30 to 100 tasks from production work. Hold the container, prompt, tools, and acceptance tests constant. Track pass rate, tool calls, input and output tokens, P50 and P95 latency, retries, human rework, and total cost per accepted task.

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

Muse Spark 1.3 earns a migration when the claimed 20% tool-call and 25% token reductions survive that evaluation. A one-point public benchmark advantage has little deployment value if it also increases latency, refusals, or review time.

#### Who Should Test It First?

Teams already using Muse Code or the Meta Model API have the clearest test path. Their workloads should include repository-wide changes, long plans, multiple tools, and detailed constraints. Version 1.2 provides a direct baseline with fewer environmental differences.

Services that perform short completions, summaries, or single-file edits do not need to migrate because of a 75.4% score. Teams waiting for a public reproduction, general `max` availability, or downloadable weights can also keep their current model.

Meta says a Muse Spark open-weights release is on its roadmap. The company did not provide a date or confirm that the future release will be the same 1.3 checkpoint.

#### Frequently Asked Questions

##### Is Muse Spark 1.3 available now?

Yes. Meta says the model is rolling out in Muse Code and the Meta Model API. Existing reasoning modes are available, while `max` will follow after additional safety testing.

##### Is Muse Spark 1.3 first on DeepSWE?

Meta's scorecard reports 75.4%, above the 74% leaders on DeepSWE's public board on September 4. The public board does not yet list Muse Spark 1.3, so the precise description is a Meta-run result above the current leaders, not a verified public-board win.

##### Can the 75.4% score be compared directly with the public board?

Both use mini-swe-agent, which makes the comparison closer than a cross-harness test. Meta still ran version 1.3 itself and pulled the other scores from the leaderboard. Its card also omits an uncertainty range for 1.3. Teams should wait for a public listing or reproduce the run before treating a narrow lead as decisive.

##### Can developers download Muse Spark 1.3 weights?

Current official access is through Muse Code and the Meta Model API. Meta announced a future Muse Spark open-weights release but did not provide a date or say that it will use the same 1.3 checkpoint.

#### Sources

*   [Meta AI Research - Introducing Muse Spark 1.3](https://research.meta.ai/blog/introducing-muse-spark-1-3)
    
*   [Meta AI Research - Muse Spark 1.3 Evaluation Methodology](https://research.meta.ai/static/muse-spark-1-3-multimodal-evaluation-methodology)
    
*   [Datacurve - DeepSWE v1.1 Public Leaderboard](https://deepswe.datacurve.ai/)
    
*   [Scale Labs - SWE-Atlas Codebase QnA](https://labs.scale.com/leaderboard/sweatlas-qna)
    
*   [Terminal-Bench - Verified 2.1 Leaderboard](https://www.tbench.ai/leaderboard/terminal-bench/2.1?verified=true)
    

#### Author Insight

The most important signal in Muse Spark 1.3 is not the 1.4-point gap between 75.4 and 74. Model vendors now compete with a coupled model, reasoning mode, and agent system. That is closer to real development, but it also makes the phrase "best model" incomplete. Engineering teams need the best reproducible system for their repository, controls, and budget.