Meta brings Muse Spark Open to the desk. Bring plenty of memory.

The model comes to your infrastructure

Meta’s release includes downloadable weights—the learned numerical settings that make the model work—and access through outside API providers. Self-hosting means running those weights on infrastructure you manage. That gives you control over where inference happens; it also makes deployment and operating the service your responsibility.

The model card lists 816 billion parameters, with 42 billion active for each generated token. This mixture-of-experts design selects part of the model for each step, reducing computation without making the remaining weights disappear from memory. It accepts text, images, video and audio, produces text, and supports a context window of one million tokens.

This is a separate access proposition from Muse Wonder’s hosted preview, which we examined earlier. Open weights let teams choose their own serving environment. They do not automatically promise the same capability, speed or commercial terms as Meta’s other models.

A desk, with infrastructure attached

Meta’s deployment guide specifies two M5 Ultra Mac Studios with 512GB of unified memory each, or four with 256GB each, connected through Thunderbolt 5 and RDMA. The latter lets the machines exchange data with low overhead. These are documented configurations, not a claim that any pair of Macs will do.

Apple’s own benchmark explains the memory issue: even the compressed model leaves too little room on one machine once intermediate calculations are included. In its NVFP4 test, two M5 Ultra systems using tensor parallelism generated 24.99 tokens per second; four reached 36.73. The context was 8,192 tokens and the measured output 128 tokens. Apple is a launch partner selling the hardware, and this is its benchmark, not our test.

A token-generation rate is only part of the wait. Reading the initial material, reasoning and making tool calls also take time. Nor does an 8,192-token benchmark establish performance at the model’s full million-token context. Before ordering machines, a team needs to test its own documents, concurrency and tolerable response time.

A strong US result, with a narrower agent story

Artificial Analysis, which received access before release, reports an Intelligence Index score of 39 at the model’s “xhigh” reasoning setting. Its launch evaluation places Spark Open first among US open-weight models, but six points behind GLM-5.3 at “max”. Geography narrows that headline considerably.

Its more revealing finding concerns the work Meta emphasises. Among the nine highest-scoring open-weight models in that evaluation, Spark Open came last on the index’s four agentic tests, but fourth on the other six combined. That is evidence about a particular suite and settings, not proof that it will fail every agent task. The evaluator’s model page identifies the component tests and links to their methods.

The useful detail is in the mistakes

Analytics software company Plotly’s first-look report tests 43 questions on a synthetic wind-farm warehouse, keeping the agent, tools and prompts constant. Spark Open passed 38, tied with Qwen 3.8 27B and below Meta’s hosted models.

It solved all 20 analytics questions but found only two of five hidden patterns, mislabelled a valid value as defective and missed a possible seasonal confounder in a correlation. A plausible answer still needs inspection.

Plotly sells the tested software and keeps its dataset private, preventing our independent reproduction. Its OpenRouter run establishes neither local Mac speed nor the quality of other compressed checkpoints.

Read the conditions before calling it free

The Muse Model License permits use and modification subject to its terms and usage policy. It requires “Built with Muse Spark Open” attribution for products or services using the model or derivatives, allowing placement in documentation, a website or settings as well as the interface.

Its revenue condition is specific: a business providing model-as-a-service must notify Meta and obtain a separate agreement before offering Spark Open or derivatives if the licensee and affiliates exceed $20 million aggregate gross revenue during any consecutive 12-month period. The clause exempts internal use, research and products that use the model as a component without themselves being model-as-a-service.

For a European team, the concrete opportunity is control over the serving environment. That still needs a deployment plan: where files are stored, which connected tools send information elsewhere, how access is managed, and how outputs are checked. A local model is one part of that system.

The release earns a serious trial. The next decision is whether the control is worth the hardware, operating work and response time for a specific job. The outside tests give that trial better starting questions than the launch slogan alone.