Meta returns to open weights with Muse Spark Open—but the license has conditions

What Meta released

Meta released Muse Spark Open on October 15 for coding and multistep agent tasks. The multimodal model has 816 billion parameters, activates 42 billion per token and supports a one-million-token context. Its weights are downloadable, with hosted access also available through providers including Fireworks and Together AI. Local operation requires substantial hardware: Meta demonstrates two M5 Ultra Mac Studios, each with 512 GB of unified memory. That is the company’s desktop configuration, rather than a promise that an ordinary laptop can run it.

The comparison mixes several kinds of evidence

Meta’s evaluation methodology combines official leaderboards, external evaluators, internal runs and competitors’ self-reported results. It therefore cannot be described simply as an entirely internal test. JobBench covers 65 professional tasks across 35 occupations. MCP-Atlas’s pass rate is the share of tasks reaching 0.75 mean claim coverage. Coding tests also use different agent setups for different models. Meta warns that its own tools and prompts may not deliver competitors’ best performance. Readers should inspect each benchmark’s method before treating a broad lead as proof for their own workload.

Read the commercial conditions before building

The Muse Model License allows use and modification subject to conditions. Copies must carry the copyright notice, license and usage policy; products need the “Built with Muse Spark Open” attribution. A model-as-a-service provider whose group gross revenue exceeds $20 million in any consecutive twelve months must notify Meta and secure a separate agreement before offering Muse or its derivatives in that business. The clause exempts internal use, research and ordinary product features outside model-as-a-service. Downloadable weights therefore do not remove contractual obligations.