Meta’s Muse Spark leans on multi-agent parallel reasoning

On October 3, 2026, a report from Juejin said that Meta had released a model called Muse Spark, built by Meta's Superintelligence Labs over a nine-month development effort. The report frames the model's central bet as a multi-agent parallel reasoning architecture, in which multiple agents work through a problem in parallel rather than along a single serial reasoning path.

According to the report, Muse Spark is positioned against GPT-5.4 Pro and Gemini 3.1 Deep Think, putting it in the same tier as those two models. On FrontierScience, described in the report as a science reasoning test, the report says Muse Spark comes out ahead. The available excerpt, however, is cut off mid-sentence at the point where the margin would be quantified, so the lead can be confirmed only as a claim made in the report, not as a measured gap.

That is the full extent of what the material contains: an October 3, 2026 date, a nine-month development timeline, a multi-agent parallel reasoning architecture, two comparison models, and a stated lead on one science reasoning benchmark. No parameter counts, context window, license or availability terms, pricing, or release schedule appear anywhere in the material, and no independent measurements or third-party reproductions are offered. The report also gives no detail on how the multi-agent setup is implemented or how the parallel agents are coordinated, and it does not say whether Meta has published a technical report, a model card, or API access. The source is a single report, so the claims about performance against GPT-5.4 Pro and Gemini 3.1 Deep Think remain unverified by anything else in the material.