🔍 Read the full analysis: Can Muse Spark Open Compete At The Price Of Meta’s Closed Model? on ThorstenMeyerAI.com
Get smart everyday buys 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 October 15, 2026, making the 816-billion-parameter model’s weights available on Hugging Face. Artificial Analysis scored it 38.9, with particularly strong results on its hallucination and visual reasoning measures, but its median provider prices match closed Muse Spark 1.3 while its overall score is lower. The source says the available pricing and license conditions complicate comparisons with cheaper, permissively licensed competitors.
Meta released Muse Spark Open, an 816-billion-parameter model with weights available on Hugging Face, on October 15. Artificial Analysis scored it 38.9 on its Intelligence Index, while the median listed provider prices match those of Meta’s closed Muse Spark 1.3, which scored higher. The release gives developers access to a large open-weight model, but its performance, price and license terms raise questions about when that access is worth the cost.
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 video as input, with text output. Weights are offered in BF16 and NVFP4 formats. Meta’s model card estimates self-hosting requires about 700 GB of GPU memory for NVFP4 or 2 TB for BF16, placing local deployment beyond the capacity of many individual developers.
Artificial Analysis tested the model at its default xhigh reasoning effort on the Intelligence Index version 4.3.2. Its score of 38.9 rounds to 39. The source account describes it as the highest-scoring open-weight model from a US lab and tenth overall in the open-weight rankings at the time of testing; the nine models ahead were from Chinese labs. Those rankings reflect the index and its tested entries, rather than every model that may be available.
The four third-party providers listed in the source 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. Those are the same listed token prices as closed Muse Spark 1.3. Artificial Analysis scored that closed model at 45.1 under the same reasoning effort, roughly six points above Muse Spark Open. The weights themselves are available without a purchase, but using a hosted provider incurs token charges.
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
What Open Access Costs
The release gives developers and organizations a way to inspect and run weights from Meta’s flagship Muse Spark line, rather than relying only on a closed service. That can matter for teams that need control over where a model runs, how it is adapted, or how data is handled. But the model’s stated memory requirements make self-hosting a substantial infrastructure undertaking, and hosted access carries the same listed token rates as the closed Muse Spark 1.3.
Price comparisons also depend on the task. Artificial Analysis put the cost of a weighted Intelligence Index task at $0.84 for Muse Spark Open. The source lists MiMo-V2.6-Pro at 46.3 for $0.13 per task and DeepSeek V4.1 Flash at 39.5 for $0.27. These figures suggest that the open label alone does not make Muse Spark Open a low-cost choice for hosted use. They are benchmark estimates, not a forecast of any particular user’s bill.
The reported results point to a more specific case for the model: document and research work that benefits from long context, visual reasoning and fewer confident wrong answers. Artificial Analysis recorded a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison, and an 81% result on its visual reasoning test. The relevance of those scores depends on how closely a developer’s actual workload matches the benchmarks.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Line
Muse Spark Open is the first open-weight model in Meta’s flagship Muse Spark line, according to the source. Meta’s recent open releases were smaller: Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the same current index. Llama 4 Maverick, released in April 2025, scored 9. The new model marks a large increase in the scale and benchmark performance of Meta’s open releases.
That improvement does not put it at the top of the overall open-weight field. The source identifies Xiaomi’s MIT-licensed MiMo-V2.6-Pro as the index leader at 46.3, with several other models also scoring above Muse Spark Open. NVIDIA’s Nemotron 3.5 Super VL, released on the same day, scored 26.5. The comparison is relevant to Meta’s position among US open-weight models, but licensing differs between these systems and affects how they can be used.
The model’s results vary by task. Artificial Analysis reported 89.2 on MCP-Atlas and 80.4 on OSWorld Verified, while its result on the more difficult Terminal-Bench 4.0 was 8%. The source also reports 76% on Terminal-Bench 2.1, a different test version. These results should not be combined into a single claim about coding or agent performance: they measure different tasks, and the lower Terminal-Bench 4.0 result contrasts with some of the stronger agent-related scores.
““built for coding and agentic work””
— Meta, in the Muse Spark Open model card
high memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Workload Fit
The source material does not provide the full text of Muse Spark Open’s license, so its precise conditions for commercial use, redistribution and other deployments cannot be established here. It describes the license as having conditions and contrasts it with MIT-licensed competitors. Developers should check Meta’s license and model card directly before deciding whether a particular use is allowed.
It is also unclear how the listed provider prices and benchmark costs will compare with real production workloads. Actual bills depend on input and output volume, caching, provider rates and task design. In Artificial Analysis’s index run, input accounted for $0.67 of the reported $0.84 task cost, including cache reads and writes. That is a benchmark-specific breakdown, not a guaranteed cost split for other uses.
Finally, the rankings describe a particular test and set of available entries. They do not establish how the model will perform in a given organization’s applications, or whether the reported results will hold as providers update deployments and prices. The source gives no independent evidence about the model’s performance outside the reported evaluations.
As an affiliate, we earn on qualifying purchases.
Check Terms Before Deployment
The immediate next step for developers is to review Meta’s model card and license, then compare provider prices against the expected mix of input, cached input and output for their own tasks. Teams considering self-hosting can use Meta’s memory estimates to assess whether NVFP4 or BF16 deployment fits their available hardware. Any benchmark comparison should account for the test version and reasoning setting used.
Further provider listings, independent evaluations and real-world deployment reports may clarify how Muse Spark Open performs on document analysis, visual reasoning and agent tasks outside the benchmark suite. The source does not specify a date for additional tests, license changes or new pricing. Until those details emerge, the release establishes the model’s current scores and listed costs, but not its value for every workload.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released Muse Spark Open on October 15, 2026, with weights available on Hugging Face, according to the source.
How did Muse Spark Open score?
Artificial Analysis scored it 38.9 on its Intelligence Index version 4.3.2 at xhigh reasoning effort. The source places it tenth among open-weight models in that ranking at the time of the test.
Is Muse Spark Open cheaper to use than Muse Spark 1.3?
Not at the listed hosted token rates in the source: the median prices from four providers matched those for closed Muse Spark 1.3. The weights are available without a purchase, but self-hosting requires substantial memory, which Meta estimates at 700 GB for NVFP4 or 2 TB for BF16.
What tasks appear strongest in the reported results?
Artificial Analysis reported an 81% result on MMMU-Pro, its visual reasoning test, and a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison. These are benchmark results and may not predict performance on a particular workload.
What should developers verify before using it?
Developers should read Meta’s current license and model card to confirm usage conditions, then check provider prices and hardware needs against their intended deployment. The source describes the license as conditional but does not include its full terms.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
