🔍 Read the full analysis: Muse Spark Open Offers Strong Open-Weight AI At A Familiar Price on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- 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 measured a score of 38.9, strong visual reasoning and a 24% hallucination rate, but found weak performance on a harder terminal-agent benchmark and a $0.84 cost per index task.
Meta released Muse Spark Open on October 15, making the weights for an 816-billion-parameter mixture-of-experts model available on Hugging Face. Artificial Analysis gave it a 38.9 score on its Intelligence Index, while its tests also found strong visual reasoning and a relatively low hallucination rate; the release gives developers a powerful open-weight option, though serving costs and the model’s mixed agent results may limit where it fits.
Muse Spark Open has 42 billion active parameters per token and 256 experts, with eight active for each token. It accepts text, images, audio and video, produces text, and has a stated one-million-token context window. Meta 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, 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 and reported 38.9 on Intelligence Index v4.3.2, rounded to 39 in its ranking. The source material identifies it as the highest-scoring open-weight model from a US lab and tenth overall at the time of testing. The nine higher-ranked open-weight models were from Chinese labs. Those rankings are tied to this benchmark and test date, not a general measure of performance on every task.
Access through four third-party providers had a median listed 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 match the list prices for Meta’s closed Muse Spark 1.3. Weights are available without a charge, but hosted use is billed by token. Artificial Analysis estimated a $0.84 cost per Intelligence Index task, with $0.67 attributed to input tokens.
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 Model With a Real Cost Trade-Off
The release changes the scale of Meta’s open-weight offerings. Muse Spark Open is the first open-weight model in the Muse Spark flagship line, and its index score is well above the smaller Meta models cited in the source: Muse Glimmer scored 17 and Llama 4 Maverick scored 9 on the same current index. That gives researchers and companies an open-weight model from a US lab that can be run on their own infrastructure, subject to the applicable license and substantial hardware needs.
Its most compelling measured results are not a clean win across all uses. 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 MMMU-Pro, leading the 66 open-weight entries tested. These results, alongside the long context window, could make it useful for visual analysis and document-heavy research. They do not establish that it will be more reliable in every production setting.
Cost and capability complicate the case for choosing it. The closed Muse Spark 1.3 scored 45.1 at the same listed token prices, according to the supplied comparison. Among open-weight options, Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3 and was listed at $0.13 per index task; DeepSeek V4.1 Flash scored 39.5 at $0.27. The comparison suggests that a high score or open weights alone do not make Muse Spark Open the lowest-cost choice. Actual expenses will depend on deployment, usage patterns and provider pricing.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Line Gets Larger
Meta’s earlier open releases cited in the source were smaller. Muse Glimmer, a 30-billion-parameter model, was released in August 2026 under Apache 2.0 and scored 17 on the current Artificial Analysis index. Llama 4 Maverick, released in April 2025, scored 9. Muse Spark Open therefore marks a substantial jump in model scale and benchmark performance within Meta’s open-weight portfolio.
The score does not put it at the top of the overall open-weight ranking. Artificial Analysis listed MiMo-V2.6-Pro at 46.3, with several other models from Chinese labs also ahead. The supplied material also contrasts Muse Spark Open with NVIDIA’s Nemotron 3.5 Super VL, released the same day, which scored 26.5. These comparisons reflect specific Artificial Analysis evaluations; different benchmarks and deployment requirements can produce different rankings.
The model’s agent results illustrate why task-level comparisons matter. Meta’s model card describes it as built for coding and agentic work, and reports results including 89.2 on MCP-Atlas and 80.4 on OSWorld Verified. Artificial Analysis recorded 76% on Terminal-Bench 2.1 but only 8% on Terminal-Bench 4.0, the harder version used in its Intelligence Index. The source also reports lower results than several peers on AutomationBench-AA and GDPval-AA. The tests point to uneven performance, rather than a simple verdict that the model is or is not suitable for agents.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card
As an affiliate, we earn on qualifying purchases.
Licensing, Deployment and Use-Case Limits
The source material does not provide the full license text or explain which conditions apply to commercial use, redistribution or fine-tuning. The weights are available on Hugging Face, but availability should not be treated as proof of unrestricted use; developers need to review Meta’s license and model card before deployment.
Reported prices are median list prices from four third-party providers, not a guarantee of what every user will pay. The $0.84 figure is an Artificial Analysis estimate for its benchmark task, and the supplied account says that 80% of the cost came from input tokens. Real-world bills may differ with prompt length, caching, output volume, provider rates and workload.
The benchmark results do not settle how the model performs across live applications, nor do they establish the reliability or safety of autonomous agent use. The supplied material also says that Muse Spark Open’s open-model score trails several competitors, while describing especially strong results on selected visual and hallucination measures. Independent testing beyond these reported evaluations will be needed to establish how those strengths hold up in specific deployments.
As an affiliate, we earn on qualifying purchases.
What Developers Should Check Next
Developers considering the model can access its weights through Hugging Face and compare local deployment against hosted inference. Before integrating it, they will need to check the license, confirm hardware and software compatibility, and test their own workloads rather than relying on a single overall index score.
For teams evaluating document or visual research tasks, the reported MMMU-Pro, long-context and hallucination results offer specific benchmarks to investigate. Teams planning terminal-based agents should pay particular attention to the 8% Terminal-Bench 4.0 result and run controlled tests before committing to unattended use. The supplied source gives no announcement of a future model update or additional release date, so the next evidence will come from further testing, provider availability and any clarification Meta makes about licensing or support.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
Muse Spark Open is Meta’s 816-billion-parameter mixture-of-experts model, with 42 billion parameters active per token. Meta released its weights on Hugging Face on October 15, 2026.
How did it score against other open-weight models?
Artificial Analysis reported a 38.9 Intelligence Index score at xhigh reasoning effort. The source ranks it tenth among open-weight models overall at the time of testing and says it was the highest-scoring model from a US lab.
Is the model free to use?
The weights are available on Hugging Face, but the source does not give the full license terms. Hosted access costs a median listed $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens through the four providers cited.
What are its strongest reported results?
Artificial Analysis reported a 24% hallucination rate and an 81% score on MMMU-Pro, leading the open-weight entries tested for that visual-reasoning benchmark. Those results apply to the cited evaluations and do not guarantee performance on every task.
Is it a strong choice for terminal-based agents?
The results are mixed. Artificial Analysis measured 8% on Terminal-Bench 4.0, despite a 76% result on the less difficult Terminal-Bench 2.1. Teams should test the model on their own agent tasks before using it for unattended shell work.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
