🔍 Read the full analysis: The Pricing Story Behind Muse Spark Open, An Open-Weight AI Model on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Spark Open on October 15, 2026, with its weights available on Hugging Face. Artificial Analysis scored it 38.9, citing strong visual reasoning and a low hallucination rate among leading open-weight models, but the model costs $0.84 per index task and performed poorly on a harder terminal-agent benchmark. Its license has conditions, and self-hosting requires substantial memory.
Meta released Muse Spark Open on October 15, making the 816-billion-parameter model’s weights available on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, while provider pricing matches Meta’s closed Muse Spark 1.3; the release therefore gives developers access to the weights, but does not make hosted use cheaper.
Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token and 256 experts, eight active at a time. It accepts text, images, speech and video, produces text, and has a one-million-token context window. Meta lists BF16 and NVFP4 weights. Its model card estimates self-hosting needs of 2 terabytes of GPU memory for BF16 or 700 gigabytes for NVFP4, and says the model is optimized to run across 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 tenth among open-weight models in the index and the highest score for an open-weight model from a US lab, according to the supplied source. The nine models above it are from Chinese labs. Within Meta’s own lineup, closed Muse Spark 1.3 scored 45.1 at the same reasoning effort, while Muse Wonder scored 52.4.
Four third-party providers list median prices of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. The source reports these prices match Muse Spark 1.3’s list prices. Weights are available without a purchase price, but running the model through a provider still 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
Strong Reasoning, Costly Runs
The release expands access to a large model from Meta’s flagship Muse Spark line, and its results stand out in specific areas. Artificial Analysis recorded an AA-Omniscience score of 24.7, the highest among 167 open-weight entries it tested. The measure rewards correct responses and penalizes confident wrong answers. The model also scored 81% on MMMU-Pro, first among the 66 open-weight models tested on that visual reasoning benchmark.
The results may appeal to teams handling long documents, research material or visual inputs. The reported hallucination rate was 24%, the lowest among the 15 highest-scoring open-weight models in the comparison. The source also reports 89% on GPQA Diamond, 44% on Humanity’s Last Exam and 78% on AA-LCR long-context reasoning. These figures describe particular evaluations; they do not establish how the model will perform on every user’s workload.
Cost changes the practical comparison. Artificial Analysis priced one Intelligence Index task at $0.84 for Muse Spark Open. MiMo-V2.6-Pro scored 46.3 for $0.13 per task, while DeepSeek V4.1 Flash scored 39.5 for $0.27. The source says the Muse task cost includes $0.67 in input charges, with cache reads accounting for $0.40. For repeated or long-running workloads, both model choice and prompt design can affect the bill.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Line
Muse Spark Open is the first open-weight release from Meta’s Muse Spark model line. The company’s recent open models were smaller: Muse Glimmer, released in August 2026 under Apache 2.0, has 30 billion parameters and scored 17 on the current index. Llama 4 Maverick, released in April 2025, scored 9, according to the source.
The release alters the size comparison for US open-weight models, but not the overall ranking reported here. NVIDIA’s Nemotron 3.5 Super VL, released on the same day, scored 26.5. The source describes it as the leading US model among permissively licensed models; Muse Spark Open scored higher but has a license with conditions. Across the broader open-weight group, Xiaomi’s MIT-licensed MiMo-V2.6-Pro led the cited index results at 46.3.
The benchmark picture is mixed. Meta’s model card describes Muse Spark Open as built for coding and agentic work. The source reports scores of 89.2 on MCP-Atlas, 80.4 on OSWorld Verified and 76.4% on Terminal-Bench 2.1. On the harder Terminal-Bench 4.0 evaluation, however, Artificial Analysis measured 8%. It reported 33% for Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3.
““built for coding and agentic work””
— Meta’s model card
As an affiliate, we earn on qualifying purchases.
License and Workload Limits
The source describes Muse Spark Open’s license as having conditions but does not provide its terms. Developers should review the license itself to determine what uses and redistribution it permits. The source also does not specify the hardware configuration, throughput or operating costs behind Meta’s Mac Studio guidance, so the published memory estimates do not by themselves establish the cost of self-hosting.
Benchmark scores do not settle how the model will perform on a particular workload. The sharp difference between Terminal-Bench 2.1 and 4.0 shows that results depend on the evaluation. The source advises against unattended shell agents on this model, but that is the author’s assessment; individual deployments would need their own evaluation. The supplied material ends during a comparison of input-token use, so it does not provide complete figures for that comparison.
As an affiliate, we earn on qualifying purchases.
Check Terms and Test Costs
Developers considering the release can inspect the Hugging Face weights and model card, review the license conditions, and check provider prices before choosing a deployment. Teams planning agent workloads should test the tasks they expect to run, including failure and retry behavior, because the reported Terminal-Bench 4.0 result was low and index-task costs include substantial input charges.
The supplied source does not identify a later Meta release or a scheduled next benchmark update. Further provider listings, license clarification or independent evaluations could change how teams assess the model’s cost and fit.
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, with weights available on Hugging Face.
How did Muse Spark Open score?
Artificial Analysis reported a score of 38.9 on its Intelligence Index at xhigh reasoning effort. The source ranks it tenth among open-weight models in that index.
Is hosted use free because the weights are available?
No. The weights are available on Hugging Face, while four providers list token prices. The source gives median prices of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens.
What is the model’s reported hallucination rate?
Artificial Analysis reported 24%, the lowest rate among the 15 highest-scoring open-weight models in the cited comparison. This is a benchmark result, not a guarantee for other tasks.
Is Muse Spark Open strong at terminal agents?
Results vary by benchmark. The source reports 76.4% on Terminal-Bench 2.1 and 8% on the harder Terminal-Bench 4.0 evaluation, so the results do not support a single broad conclusion about agent performance.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
