🔍 Read the full analysis: Muse Spark Open Brings Strong Open Weights With A Meta-Style Price Tag 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 October 15, 2026, making the 816-billion-parameter model’s weights available on Hugging Face. Artificial Analysis scored it 38.9, with strong results on visual reasoning and a hallucination benchmark, but the hosted token prices match closed Muse Spark 1.3, and the model trails competitors on some agent tasks.
Meta released Muse Spark Open on October 15, putting the weights for an 816-billion-parameter model on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, but the model’s hosted per-token rates match Meta’s higher-scoring closed Muse Spark 1.3, making the release a significant open-weight offering with a notable price caveat.
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. It supports a one-million-token context window and accepts text, images, speech or audio, and video; its output is text. Meta has published BF16 and NVFP4 weights on Hugging Face. Its model card estimates self-hosting requires about 700 GB of GPU memory for NVFP4 or 2 TB for BF16.
Artificial Analysis tested the model at its default “xhigh” reasoning effort and reported an Intelligence Index score of 38.9, rounded to 39. The source says that places it tenth among open-weight models overall and makes it the highest-scoring open-weight model from a US lab in that comparison. The nine open-weight models ranked above it are from Chinese labs. These are benchmark results, not a guarantee of performance on a particular workload.
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. The source reports that these rates match those for closed Muse Spark 1.3. The weights are available, but using a hosted provider is not free; self-hosting also calls for substantial hardware capacity.
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
Open Weights, Closed-Model Rates
The release gives developers access to Meta’s flagship Muse Spark model weights rather than requiring use of the closed service. That can matter to organisations seeking to run a model on their own infrastructure or examine and adapt its weights, subject to the applicable license. But access does not remove the cost of computing: Meta’s stated memory requirements are high, and the hosted token rates match the closed model’s.
The benchmark picture is mixed. Artificial Analysis reports strong performance on AA-Omniscience, where Muse Spark Open scored 24.7, the highest among 167 open-weight entries in that test. It also scored 81% on visual reasoning benchmark MMMU-Pro, leading the 66 open-weight models tested there. Its reported hallucination rate was 24% among the 15 highest-scoring open-weight models, the lowest in that group. Those results could interest teams working with research or document-heavy tasks, although benchmark performance alone does not establish reliability in deployment.
For buyers comparing hosted models, the benchmark score is only part of the decision. The source estimates a weighted cost of $0.84 per Intelligence Index task. It lists MiMo-V2.6-Pro at a higher score of 46.3 for $0.13 per task and DeepSeek V4.1 Flash at 39.5 for $0.27. These comparisons depend on the source’s testing and pricing methodology; actual bills will vary with usage and provider rates.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Meta’s Open-Weight Lineup
Muse Spark Open is the first open-weight release in Meta’s Muse Spark line. The company’s recent open models were smaller: the source lists Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, at 17 on the same current index. Llama 4 Maverick, released in April 2025, is listed at 9. The new release therefore marks a substantial change in the scale of Meta’s open-weight offerings.
It does not lead the overall open-weight ranking cited by the source. Xiaomi’s MIT-licensed MiMo-V2.6-Pro scores 46.3, while several other models from Chinese labs also rank above Muse Spark Open. Within the US open-weight field, the source compares it with NVIDIA’s Nemotron 3.5 Super VL, released the same day and scored at 26.5. These standings refer to the Artificial Analysis Intelligence Index and should not be treated as a universal ordering across every task.
Another relevant comparison is Meta’s own closed Muse Spark 1.3. At the same reported reasoning effort and listed token prices, the closed model scored 45.1, compared with 38.9 for Muse Spark Open. Muse Wonder is listed at 52.4 in the source. This makes the open release’s practical appeal dependent not just on scores, but on whether access to weights, deployment control or other model characteristics justify its costs for a given use.
““built for coding and agentic work””
— Meta, in the Muse Spark Open model card
As an affiliate, we earn on qualifying purchases.
License and Workload Fit
The supplied source does not state the full terms of Muse Spark Open’s license, so the extent of permitted commercial use, redistribution or modification cannot be established from the available material. The model is described as open-weight, but that term does not by itself mean the weights are under a permissive or unrestricted license.
Results also vary sharply by task. The source reports 8% on Terminal-Bench 4.0, a harder terminal-agent benchmark used in the Intelligence Index, while Meta’s model card cites 76.4% on Terminal-Bench 2.1. Those are different benchmark versions and are not directly interchangeable. The same source reports 50% on AutomationBench-AA and 1,401 Elo on GDPval-AA, but does not provide enough methodological detail here to predict results for a specific application.
Pricing figures are median list prices from four third-party providers, not a guaranteed rate for every provider or customer. The source’s account of per-task costs also describes a particular benchmark run. Real costs depend on token volumes, caching, provider pricing and task length. The hardware estimate is from Meta’s model card; the material does not establish how performance or operating costs compare across different self-hosted setups.
As an affiliate, we earn on qualifying purchases.
Deployment Choices Ahead
Developers can access the published BF16 and NVFP4 weights on Hugging Face and review Meta’s model card and license before deciding whether to self-host or use a third-party provider. Teams evaluating the model will need to test it on their own workloads, especially if they plan to use it for coding, long-running agents, visual reasoning or large document sets.
Further provider listings, licensing details and independent testing could clarify how the model performs and what it costs outside the reported benchmarks. For now, the confirmed release establishes that a large Meta Muse model is available as open weights; it does not establish that it is the cheapest or strongest choice across use cases.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released the model on October 15, 2026. Its weights are available on Hugging Face.
How large is Muse Spark Open?
It has 816 billion total parameters, with 42 billion active per token. Meta’s model card estimates 700 GB of GPU memory for NVFP4 self-hosting and 2 TB for BF16.
How did it score against other open-weight models?
Artificial Analysis scored it 38.9 on its Intelligence Index v4.3.2. The source ranks it tenth among open-weight models overall and says nine models, all from Chinese labs, scored higher.
Is Muse Spark Open cheaper than Meta’s closed model?
Not at the hosted list prices reported in the source. Four third-party providers’ median rates match those of closed Muse Spark 1.3: $1.25 per million input tokens, $0.15 for cached input and $4.25 for output. The weights can be downloaded, but self-hosting requires substantial hardware.
Is it suited to coding agents?
Meta’s model card describes it as built for coding and agentic work, and the source reports positive results on some tests. However, Artificial Analysis measured 8% on Terminal-Bench 4.0. Results depend on the benchmark and task, so teams should test the model against their own requirements.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
