AGTP (@AGTPinsights) on X

What it claims

An X thread from AGTP summarizes Meta's Muse Spark 1.3 Evaluation Report. Its central claim comes from Exponential Security Labs: Muse Spark 1.3 is about as robust to prompt injection as GPT-5.6 Sol. The thread calls Muse Spark 1.3 the first model Meta released in the Muse family. It shipped on September 2, 2026, with gains in agentic and coding tasks. Zifan (Sail) Wang announced the report in a separate thread.

Evidence

• Prompt injection. Meta used two new Exponential benchmarks, RepoInject v1 (repositories) and BrowserInject v1 (browsing). It also ran Gray Swan's IPI Arena on the Q1 and Q2 splits, which run through April, as other labs' model cards did. Meta reports "strong results" there, but the thread gives no scores for any of the three benchmarks.
• Chain-of-thought controllability. This puts Meta alongside OpenAI, Anthropic and Google DeepMind in reporting these evals. Researcher Tomek Korbak says controllability stays low, which suggests a weak ability to evade CoT monitors.
• Loss of Control. PostTrainBench is listed as a key catastrophic-risk eval, with no result given.
• Human red teaming. Handshake red-teamed the model before launch (report section 3.1.4.2). It had pre-release access and found jailbreaks through the product. It sent full trajectories back to Meta so mitigations could improve.
• Capability numbers. The thread lists 75.4% on DeepSWE 1.1, 98.5% on long-context MRCR and a 1M-token context window. It gives no baseline or settings for these.

How to use it

Muse Spark Open is Meta's open-weight model. It has 816B total parameters, 42B active, and a 1M-token context window. Applied Compute has it live on AC2 for training and inference. Together AI hosted a live session with Meta about it. The thread does not state a licence.

Limits

• Prompt injection is not solved. Maksym Andriushchenko of Exponential says so, despite the GPT-5.6 Sol comparison.
• Compliance is not reported. Korbak says he wishes Meta had reported compliance alongside controllability, so that half of the CoT picture is missing.
• Muse Wonder has no safety report. The Midas Project says Meta released Muse Wonder on October 12, 2026, with only a 13-page model card and a promise of a safety report soon. It says Meta updated the card 11 days later, and there is still no safety report. Engineers evaluating Wonder have no safety write-up to compare with Spark 1.3.
• The numbers are hard to attribute. The thread does not say whether the DeepSWE 1.1 and MRCR figures belong to Muse Spark 1.3 or to Muse Spark Open. Check the report itself before citing them.
• The source is secondhand. This is a thread summary, not the report, and it contains no injection success rates or eval protocols.