🔍 Read the full analysis: Exploring Muse Spark Open, America’s Strongest Open-Weight Model 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 weights on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, making it the highest-scoring open-weight model from a US lab in the test, while nine open-weight models from Chinese labs scored higher. Its low measured hallucination rate and visual reasoning results stand out, but agent performance and per-task costs vary sharply.
Meta released Muse Spark Open on October 15, putting the weights for an 816-billion-parameter mixture-of-experts model on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, the highest result among open-weight models from US labs in that index, while nine open-weight models from Chinese labs scored higher overall.
Muse Spark Open has 816 billion total parameters, with 42 billion active for each token. Its configuration uses 256 experts, activating eight per token, and supports a one-million-token context window. Meta lists text, images, speech and audio, and video as inputs; the model produces text. The weights are available in BF16 and NVFP4 formats.
Artificial Analysis tested the model on release day at its default xhigh reasoning effort. Its Intelligence Index v4.3.2 score is 38.9, rounded to 39 in the published rankings, placing it tenth among open-weight models overall. The top nine entries are from Chinese labs. The index score is a benchmark result, not a measure of performance on every task.
Meta says self-hosting requires about 700 GB of GPU memory for NVFP4 or 2 TB for BF16. It also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX. Artificial Analysis listed four third-party API providers, with median prices of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output 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 Larger US Open-Weight Contender
The release changes the scale of Meta’s open-weight offerings. Muse Spark Open is the first open-weight model in its flagship Muse Spark line, and its 38.9 index score is well above the US open-weight models cited in the source: Muse Glimmer scored 17, while Llama 4 Maverick scored 9. NVIDIA’s Nemotron 3.5 Super VL, released the same day, scored 26.5. These are comparisons within the Artificial Analysis index, not a guarantee that Muse Spark Open is better for every use.
Its strongest reported results suggest potential for research and document work. Artificial Analysis recorded 24.7 on AA-Omniscience, the highest score among 167 open-weight entries in that test, and a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison. It also scored 81% on MMMU-Pro, a visual reasoning benchmark, leading the 66 open-weight models tested. The figures provide evidence for specific benchmark strengths; real-world accuracy will depend on the task and deployment.
Access to the weights does not make the model inexpensive to operate. The median listed API rates match those for Meta’s closed Muse Spark 1.3, while a weighted Artificial Analysis index task cost $0.84. That cost and the model’s hardware needs matter to teams weighing self-hosting against an API or a smaller model.
As an affiliate, we earn on qualifying purchases.
How It Compares With Alternatives
The gap between leading US and Chinese open-weight results remains visible in the supplied rankings. Muse Spark Open is tenth overall, behind nine models from Chinese labs. Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3, while DeepSeek V4.1 Flash scored 39.5. The results show that a large release from a US lab has moved the size comparison, but has not taken the top index position.
Its closest comparisons depend on what is being measured. Against Meta’s closed Muse Spark 1.3, the open model scored 38.9 versus 45.1 at the same xhigh effort, according to the source material; Muse Wonder scored 52.4. The source also reports an August score of 39.6 for Muse Spark 1.2 at xhigh effort. Those figures frame the new model’s place within Meta’s lineup, though benchmark scores alone do not settle deployment choices.
Artificial Analysis reported mixed results across agent benchmarks. Muse Spark Open scored 76% on Terminal-Bench 2.1, but 8% on Terminal-Bench 4.0, the harder version used in the Intelligence Index. On Terminal-Bench 4.0, Muse Spark 1.3 scored 33%, MiMo-V2.6-Pro 35% and GLM-5.3 42%. Other reported figures include 50% on AutomationBench-AA and 1,401 Elo on GDPval-AA. The source presents these as test results; they do not establish how the model will perform in every agent setup.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card
As an affiliate, we earn on qualifying purchases.
Deployment Costs and Agent Results
The available results do not establish how Muse Spark Open will perform across production workloads. Artificial Analysis tested it on release day, but the supplied material does not give details about independent replications, longer-term use, or how results may change with different prompts and configurations. The reported hallucination rate is specific to the evaluation and comparison group described.
Actual self-hosting costs remain unclear. Meta’s memory estimates describe hardware requirements, not the cost of acquiring, powering or operating that hardware. API pricing is based on four third-party providers and may change. The supplied source also ends partway through its discussion of input-token usage, so it does not provide a complete comparison of token consumption across models.
The low Terminal-Bench 4.0 result sits alongside higher scores on other agent tests. The source offers no confirmed explanation for that difference. Teams considering agent use would need task-specific evaluation before relying on the model for unattended work.
High-performance AI inference hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing Before Choosing a Provider
The next step for teams evaluating Muse Spark Open is to compare it on their own workloads, including the visual, long-context or research tasks suggested by its stronger reported results. Agent use calls for particular scrutiny given the gap between its Terminal-Bench 2.1 and 4.0 scores. Buyers should also compare provider pricing and estimate hardware and operating costs for self-hosting.
Artificial Analysis’ release-day measurements provide an initial reference point, while Meta’s model card gives the stated formats and hardware guidance. The supplied information does not identify a later benchmark release or a scheduled model update, so further developments remain unannounced.
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.
How did Muse Spark Open score?
Artificial Analysis measured a score of 38.9 on the Intelligence Index v4.3.2, rounded to 39. It ranked tenth among open-weight models overall and first among the US open-weight models described in the source.
Is Muse Spark Open free to run?
The weights are available, but running the model requires substantial hardware or a paid API. Meta estimates 700 GB of GPU memory for NVFP4 and 2 TB for BF16; the reported median API rates are $1.25 per million input tokens and $4.25 per million output tokens.
Is it strong at agent tasks?
Results vary by benchmark. Artificial Analysis reported 76% on Terminal-Bench 2.1 and 8% on Terminal-Bench 4.0, so those figures do not support a uniform claim of strong agent performance.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
