Muse Spark could put Meta back on the map

Meta has launched Muse Spark, the first model from its Meta Superintelligence Labs.

Meta says the model is competitive with OpenAI, Anthropic, and Google on several benchmarks, though it does not lead across the board.

If those results hold up, Muse Spark could help Meta recover after Llama 4 was poorly received.

There are still caveats. Meta has previously been criticised for using benchmark results that made its models look stronger than the versions available to users.

Muse Spark is also far more closed than Meta’s earlier models, with access mostly limited to Meta’s own apps and tools.

The model powers Meta AI and will soon roll out across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses.

It is also Meta’s first reasoning model, meaning it can work through problems step by step. It can handle text and images, use tools, and coordinate subagents.

Benchmark results show a mixed picture. Muse Spark trails rivals on advanced reasoning tests like GPQA Diamond, but outperforms them on HealthBench Hard.

Meta says it still needs to improve in coding and long-running agent tasks.

The launch is the clearest result yet of Meta’s AI overhaul after Llama 4.

Since then, the company has spent heavily on talent, infrastructure, and new leadership, including bringing in Scale AI cofounder Alexandr Wang as chief AI officer.

Meta says Muse Spark also went through extensive safety testing, though outside researchers found signs that the model could recognise when it was being evaluated.

Meta said “we’re so back” and this time it might actually be true. - MG