Muse Spark Open: Meta’s 816B Model, Hardware, License and Test

👁 116 reads

Meta’s Muse Spark Open offers downloadable weights for a large multimodal model as of October 2026, with demanding hardware requirements and its own license. I tested it via the API on October 19: it scored 11/12 in my test in Polish. The missing point came from an empty response after the token budget ran out.

What Meta released with Muse Spark Open

Muse Spark Open was released on October 15, 2026. It is a mixture-of-experts model aimed at coding, tools, and extended tasks. Its one-million-token context window describes how much context the model is designed to handle, rather than how much every installation can process efficiently. The official repository provides the weights and documentation for running it outside Meta’s own applications. It is one of four open-weight releases from that week, compared side by side in my overview of open-weight AI models in October 2026; Meta’s hosted model is covered in my Muse Spark 1.3 review.

The difference between total and active parameters matters here. The model does not perform the same computation with all 816 billion parameters for every token. It activates a smaller selection. However, the full set of experts remains part of the checkpoint, so the 42-billion active figure is not a shortcut for estimating the download or memory footprint.

The card describes multimodal inputs and text output. My own test used text prompts only. It cannot support a claim about reading screenshots, interpreting charts, or understanding recordings, even though those capabilities are part of the product’s positioning. They require separate examples and separate scoring.

The hardware requirement behind the Mac headline

According to Meta’s model card, checked October 20, the NVFP4 version requires about 700 GB of GPU memory, while BF16 requires about 2 TB. Meta also describes running the model with MLX across two M5 Ultra Mac Studio systems, each with 512 GB of unified memory. The configuration is specific enough to be useful, but it is far beyond a typical personal computer.

These are Meta’s published requirements and examples, not hardware results from my test. I did not run Muse Spark Open locally. The API result therefore says nothing about the speed of the Mac configuration, its power consumption, or how it behaves with a long context.

Memory planning should also distinguish the checkpoint from the complete running service. Conversation state, runtime buffers, and concurrent requests all affect capacity. A system that can load the weights may still need a smaller context or fewer simultaneous users than the headline specification suggests. Quantization changes that planning, but does not eliminate it. If you want a model that genuinely runs on one consumer GPU, start with my local AI guide instead.

The license has conditions that matter

Muse Spark Open uses the Muse Model License rather than MIT. The text permits use, modification, fine-tuning, and distribution subject to its conditions. It requires the wording “Built with Muse Spark Open” in the product interface or an allowed accompanying location, and it incorporates a usage policy.

A separate agreement applies to qualifying model-as-a-service businesses when aggregate revenue for the licensee and affiliates exceeds $20 million over a consecutive 12-month period. The license specifies exceptions, including internal use and product features that are not themselves model-as-a-service offerings. The relevant distinction is the actual business and use, not simply whether an application happens to call an API.

That makes the license part of deployment planning. A team evaluating internal workflows and a company selling general model access can face different conditions. The downloadable checkpoint alone does not answer whether a particular commercial arrangement fits the standard terms.

What happened in my test in Polish

On October 19, I submitted 12 short Polish prompts via the API. Every model in the comparison received the same tasks, with temperature 0 and a maximum of 6,000 completion tokens, including reasoning. Each task was run once. Muse Spark Open returned accepted answers for 11 tasks and no final answer for the remaining one. The full comparison with DeepSeek, Nemotron and Step 5 is in my test of four open models in Polish.

The failed item asked how many seconds are in 2.5 hours. The recorded request lasted 126.3 seconds, used 6,000 completion tokens, and returned an empty content field. The answer should have been 9,000. This is a failure to deliver within the configured budget; the log does not show the model confidently returning an incorrect number.

Across all 12 requests, the median elapsed time was 8.05 seconds. The recorded usage totaled 12,241 completion tokens and about $0.0542. That amount is the historical cost of this particular run via the API, not a claim about Meta’s current direct API pricing. The requests also used different serving backends, which limits what their timing can say about the model itself.

The final score includes a grading correction. The spelling task originally expected “wziąć,” but the answer “Weź parasol, bo na pewno będzie padać.” is also valid Polish. I accepted it. The correction changed the score without changing or regenerating the model’s answer.

There is another useful limit to the result: the grader tolerated some extra formatting. It checked answers and extracted JSON rather than enforcing every presentation instruction strictly. An application that requires an exact response shape should use a stricter validation rule than the one used in this small test.

Who should evaluate it

Muse Spark Open is relevant to teams that have a reason to control the weights and enough infrastructure to do so. Meta’s emphasis on extended tool use suggests evaluating it on a complete workflow, including recovery from tool errors and delivery of a final result. A short arithmetic question cannot stand in for that work.

• For internal deployment, test the actual documents and tool permissions the model will encounter.
• For fine-tuning, check the license and measure whether customization improves the intended tasks.
• For interactive use, track empty responses and long waits alongside accepted answers.
• For local deployment, budget for the running system rather than just the checkpoint.

The observed failure also argues for testing reasoning controls. A larger response allowance might let the model finish, while a lighter reasoning setting might reduce unnecessary work. Neither change was tested in this run, so neither is a demonstrated fix. Both belong in a follow-up evaluation with cost and latency recorded.

FAQ

Can Muse Spark Open run on a Mac?

Meta documents a setup using two M5 Ultra Mac Studios with 512 GB each and MLX. That should not be generalized to other Macs without checking memory and runtime support.

Does 11/12 mean it is worse at Polish than every competitor?

It completed fewer items than DeepSeek and Nemotron in this run. The sample is too small to establish a general language ranking, and the missing answer occurred at the response limit.

Is this an English-language benchmark?

No. This article is in English, but my test was in Polish. It did not evaluate English writing, translation quality, or multimodal understanding.