Meet Muse Spark Open, An Open-Weight Model Priced Like Meta’s Closed One

🔍 Read the full analysis: Meet Muse Spark Open, An Open-Weight Model Priced Like Meta’s Closed One on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

Meta released Muse Spark Open on October 15, publishing the weights for an 816-billion-parameter model on Hugging Face and opening its flagship Muse Spark line to self-hosting. Artificial Analysis scored it 38.9 on its Intelligence Index, but the model’s reported token prices are the same as those of closed Muse Spark 1.3, which scored higher at 45.1 in the same test setting.

Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token and 256 experts, eight active for each token. It accepts text, images, audio and video, produces text, and has a one-million-token context window. The Hugging Face release includes BF16 and NVFP4 weights. Meta’s model card estimates that self-hosting requires 2 terabytes of GPU memory in BF16 or 700 gigabytes in NVFP4; it also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX.

Artificial Analysis evaluated the model at its default xhigh reasoning effort and rounded its 38.9 Intelligence Index score to 39. The source reports that this is the tenth-highest score among open-weight models in the index and the highest from a US lab. The nine models ahead are from Chinese labs. The result is a substantial change in the scale of Meta’s open releases, but not a lead over the wider open-weight field.

The weights can be downloaded, but hosted access is not free. Across four third-party providers, the reported median list price is $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Those are the same list prices reported for closed Muse Spark 1.3. Artificial Analysis estimates a cost of $0.84 for a weighted Intelligence Index task on Muse Spark Open, compared with $0.13 for MiMo-V2.6-Pro and $0.27 for DeepSeek V4.1 Flash. These figures are benchmark-task estimates, not a general price for every user’s workload.

Open Weights, Closed-Model Rates

The release gives developers and organisations the option to download and self-host a model from Meta’s flagship line, rather than rely exclusively on a closed service. That can matter for teams with the infrastructure and expertise to run a very large model, or with reasons to control where inference happens. But Meta’s published memory estimates make clear that self-hosting is a substantial compute undertaking, not a low-cost default.

For customers using hosted providers, open weights do not translate into a price advantage in this case. Muse Spark Open’s reported per-token rates match the closed 1.3 model, while that closed model scored 45.1 versus 38.9 in Artificial Analysis’s evaluation at the same reasoning effort. Against some open alternatives, the gap in estimated benchmark cost is larger: the source reports MiMo-V2.6-Pro scoring 46.3 at $0.13 per index task. These comparisons can help buyers shortlist models, but benchmark costs do not predict every application’s total bill or performance.

The model has strengths that could still make it a fit for particular work. Artificial Analysis reported an AA-Omniscience score of 24.7, highest among 167 open-weight entries in that measure, and a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models. It also scored 81% on the visual-reasoning benchmark MMMU-Pro, leading the 66 open-weight entries tested there. Those results, combined with its long context window, could be relevant to document-heavy and research workflows. They do not establish that it will be more accurate or cheaper for an individual deployment.

How Muse Fits the Open Field

Muse Spark Open is Meta’s first open-weight model in the Muse Spark family. Its earlier open offerings were smaller: the source gives Muse Glimmer, a 30-billion-parameter model released under Apache 2.0 in August 2026, an Intelligence Index score of 17. It reports a score of 9 for Llama 4 Maverick, released in April 2025. The new model therefore marks a move by Meta into a far larger scale of open-weight release.

That move does not make Muse Spark Open the leading open-weight model overall. According to the supplied Artificial Analysis results, MiMo-V2.6-Pro leads at 46.3, and several other models from Chinese labs also score above Meta’s release. Among US models, the source compares it with NVIDIA’s Nemotron 3.5 Super VL, which scored 26.5 and was released on the same day. The source describes Muse Spark Open as the leading US open-weight model, while qualifying that comparison by license: it is not a permissive-license release.

Benchmarks also give a more mixed picture than the headline index score. Muse Spark Open scored 89% on GPQA Diamond and 44% on Humanity’s Last Exam, reported as fourth among open-weight models. But on Terminal-Bench 4.0, the harder terminal-agent test used in the index, Artificial Analysis measured 8%. The source reports 33% for closed Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3. Meta’s model card describes the model as built for coding and agentic work; the reported test results show that performance varies sharply by task.

“built for coding and agentic work”

— Meta, in the Muse Spark Open model card

Deployment Costs and Test Limits

The supplied material reports benchmark results and median list prices, but does not provide enough detail to determine what a specific customer will pay across providers, including any discounts, service limits or differences in implementation. The $0.84 per-task estimate reflects Artificial Analysis’s weighted benchmark run; actual costs will depend on prompt length, caching, output volume and the workload being run.

It is also not clear from the material how the model performs across a broad range of real-world deployments, or how its results compare under different inference settings. The reported benchmark does not show that the model’s low measured hallucination rate will carry over to every task. Nor does publication of weights alone establish that every organisation can use the model on its own hardware: Meta’s stated memory requirements are large, and the supplied material does not spell out the license’s full terms or restrictions.

The source also presents model-card results such as 76.4% on Terminal-Bench 2.1 alongside Artificial Analysis’s 8% on Terminal-Bench 4.0. Those are different versions of the test and should not be treated as directly comparable readings of one benchmark. More detail on methodology and how results transfer to practical work remains unavailable in the supplied account.

Provider Access and Real-World Tests

The immediate next step for users is to inspect the Hugging Face weights and model card, check the license terms, and compare self-hosting requirements with available hardware. Teams considering hosted inference can compare provider pricing and test the model on their own prompts, paying particular attention to input and cached-input charges, which accounted for most of the source’s benchmark-task estimate.

Further provider listings, independent evaluations and real-world deployment reports may clarify whether Muse Spark Open’s accuracy, visual reasoning and long-context performance justify its costs for particular applications. Developers evaluating coding or autonomous agents will also want results on the specific terminal and tool-use tasks they plan to run; the reported 8% on Terminal-Bench 4.0 is a reason to test before assigning consequential or unattended work.

For now, the confirmed development is the release of the weights and the published evaluation snapshot. The model’s competitive position, hosted economics and practical reliability will depend on additional tests and on whether users can operate it at a cost and scale that suit their needs.

Key Questions

What is Muse Spark Open?

Muse Spark Open is Meta’s 816-billion-parameter mixture-of-experts model, with 42 billion active parameters per token. Meta published BF16 and NVFP4 weights on Hugging Face on October 15, 2026.

How did it score against other open-weight models?

Artificial Analysis scored it 38.9 on its Intelligence Index at xhigh reasoning effort. The supplied results rank it tenth among open-weight models overall and first among US open-weight models, while several models from Chinese labs scored higher.

Is hosted Muse Spark Open cheaper than Meta’s closed model?

Not according to the reported list prices. Four third-party providers had a median price of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens—the same rates reported for closed Muse Spark 1.3.

Can it be self-hosted?

Meta publishes the weights and says self-hosting requires an estimated 700 GB of GPU memory for NVFP4 or 2 TB for BF16. Its model card also describes an optimized setup using two Mac Studios with M5 Ultra and MLX; the source does not establish how that setup performs for every workload.

Is Muse Spark Open suitable for coding agents?

Results vary by test. Meta describes it as built for coding and agentic work, but Artificial Analysis measured 8% on Terminal-Bench 4.0. Users considering agent deployments should evaluate it on their own tools and tasks rather than rely on the model-card description alone.

Source: ThorstenMeyerAI.com