🔍 Read the full analysis: The Case For Muse Spark Open At Meta’s Closed-Model Price on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Meta released Muse Spark Open on Oct. 15, 2026, making the 816-billion-parameter model weights available on Hugging Face. Artificial Analysis scored it 38.9, but its hosted token prices match Meta’s higher-scoring closed Muse Spark 1.3, and the license has conditions. It stands out on hallucination and visual-reasoning tests, while results and costs on some agent tasks are less competitive.
Meta released Muse Spark Open on Oct. 15, putting the weights for its 816-billion-parameter model on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, while four third-party providers list token prices matching Meta’s closed Muse Spark 1.3, which scored higher on the same index.
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. Meta lists a one-million-token context window and support for text, images, speech and audio, and video as input; its output is text. The weights are available in BF16 and NVFP4 formats. Meta’s model card estimates that self-hosting requires 700 GB of GPU memory for NVFP4 or 2 TB for BF16, and says the model is optimized to run on two Mac Studios with M5 Ultra using MLX.
Artificial Analysis tested the model at its default xhigh reasoning effort. Its score of 38.9 is rounded to 39 by the index. The source report describes it as the highest-scoring open-weight model from a U.S. lab and the tenth-highest-scoring open-weight model overall. The nine models ahead of it in that ranking are from Chinese labs, according to the report.
Four third-party providers list a median price of $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. The weights themselves are free to download, but using a hosted service is not. The report says the model’s license carries conditions; the supplied material does not specify their full scope.
Infographic 01 · Where Muse Spark Open lands
The best US open-weight model is tenth. The nine ahead are Chinese.
Artificial Analysis scores Meta’s first open Muse Spark model at 38.9. That is 12 points clear of any other US open-weight model, but behind nine Chinese models. Four of those nine are MIT-licensed, and three of them cost a third or less per task.
Top open-weight models by Intelligence Index
Higher is better. Effort variants of the same model are not listed twice. Cost = average API cost to run one index task.
License: Muse Spark Open ships under the Muse Model License: broad use and redistribution, but mandatory “Built with Muse Spark Open” attribution, a separate agreement for model-as-a-service businesses above $20M revenue, and a usage policy.
Source: Artificial Analysis, Muse Spark Open (xhigh) model page and model data, read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. License categories as listed by Artificial Analysis. Countries: Xiaomi, Z AI, StepFun, Kimi, InclusionAI, Alibaba and DeepSeek are Chinese labs; Meta and NVIDIA are US.
Powered by Thorsten Meyer AI
A large release with a costly trade-off
The release changes the scale of Meta’s open-weight offerings: Muse Spark Open is the first model from its flagship Muse Spark line to have weights released, and its 816-billion-parameter total size is far above the smaller recent Meta releases described in the source. That gives developers and researchers a new model they can download and potentially run on their own infrastructure, subject to the license and substantial hardware demands.
But availability of weights does not mean low-cost deployment. Renting access at the listed rates costs the same per token as the closed Muse Spark 1.3, even though that model scored 45.1 versus 38.9 at the same reasoning effort. On Artificial Analysis’s weighted index task, Muse Spark Open cost $0.84 in the report’s test. MiMo-V2.6-Pro scored 46.3 for $0.13, while DeepSeek V4.1 Flash scored 39.5 for $0.27. Those figures are specific to the reported test and pricing; actual bills depend on use and provider.
The model’s strongest results may matter for workloads involving documents, images and research. Artificial Analysis recorded a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison, and an 81% score on its visual-reasoning test, highest among the 66 open-weight entries tested. These are benchmark results, not guarantees of accuracy in real applications.
As an affiliate, we earn on qualifying purchases.
How it compares with other open models
Meta’s previous open-weight releases cited in the source were considerably smaller. Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the current index. Llama 4 Maverick, released in April 2025, scored 9. The source says NVIDIA’s Nemotron 3.5 Super VL, released on the same day as Muse Spark Open, scored 26.5.
Muse Spark Open’s overall score does not place it at the top of the open-weight field. The report lists MiMo-V2.6-Pro at 46.3, with several other models also scoring above Meta’s release. Its comparison also puts Muse Spark 1.3 at 45.1 and Muse Wonder at 52.4. Thus, the release is a significant increase in the scale of Meta’s open-weight models, but the benchmark evidence supplied does not show it surpassing the leading open or closed models.
Licensing is part of that comparison. The source identifies MiMo-V2.6-Pro and DeepSeek V4.1 Flash as MIT-licensed, while describing some competing models as having their own licenses. It characterizes Muse Spark Open’s license as conditional, but does not provide enough detail to compare specific permitted uses. Users should review the model’s actual license terms before deployment.
“Built for coding and agentic work”
— Meta’s model card
As an affiliate, we earn on qualifying purchases.
License limits and real-world costs
The supplied material does not set out the full terms of Meta’s license, so it is not possible from this account alone to determine which commercial, redistribution or deployment uses are allowed. The availability of downloadable weights should not be treated as proof that every use is unrestricted.
Benchmark rankings and provider list prices are snapshots, not direct measures of performance or cost in every production setting. The report gives a weighted index-task cost of $0.84, including $0.67 in input-token charges, but does not establish how a particular application’s prompts, caching, output volume or provider pricing would affect its bill. The source material also ends during its discussion of token consumption, so it does not supply a complete comparison of input use across all models.
Agent performance is mixed across tests. Artificial Analysis measured 8% on Terminal-Bench 4.0, compared with 33% for Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3. The model scored 76% on Terminal-Bench 2.1, according to the same report. The reasons for the difference between these benchmarks, and how those results translate to specific agent workflows, are not established by the supplied information.
As an affiliate, we earn on qualifying purchases.
What developers should check next
The next practical step for developers is to review the Hugging Face model card and license, then test the model on their own workloads before committing to self-hosting or a paid provider. Hardware requirements are high, and hosted prices can make repeated or long-running tasks costly. Application-specific evaluations should include accuracy checks, input and output token tracking, and tests of any agent actions that could have external effects.
Future comparisons will depend on updated benchmark results, changes in provider prices and more detail about the license’s permitted uses. For now, the published evidence supports a qualified assessment: Muse Spark Open is a large, capable model with notable results on hallucination and visual reasoning, but its hosted rates match a higher-scoring closed model and its performance on some agent benchmarks is weak.
artificial intelligence development tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is an 816-billion-parameter mixture-of-experts model released by Meta on Oct. 15, 2026. Meta says 42 billion parameters are active per token, and the weights are available on Hugging Face.
Is Muse Spark Open free to use?
The weights are available to download, but the model has high hardware requirements and its license has conditions. Third-party hosted access is paid: the reported median list prices are $1.25 per million input tokens and $4.25 per million output tokens, with a separate cached-input price.
How did it score against Meta’s closed Muse Spark 1.3?
Artificial Analysis scored Muse Spark Open at 38.9 and Muse Spark 1.3 at 45.1 at the same xhigh reasoning effort, according to the source. Their reported third-party list prices per token are the same.
What are Muse Spark Open’s strongest reported results?
Artificial Analysis recorded a 24% hallucination rate, lowest among the 15 highest-scoring open-weight models in its comparison, and an 81% score on its visual-reasoning test, highest among 66 open-weight entries tested. These results reflect benchmarks, not guaranteed performance in every use.
Is it a strong choice for coding agents?
The evidence is mixed. The source reports 76% on Terminal-Bench 2.1 but 8% on the harder Terminal-Bench 4.0 test. Developers should evaluate it on their own tasks rather than assume the model card’s description predicts results in a particular agent workflow.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
