🔍 Read the full analysis: Muse Spark Open: America’s Strongest Open-Weight Model At Meta’s Closed-Model Price on ThorstenMeyerAI.com
Get movie nights 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 Oct. 15, 2026, publishing its 816-billion-parameter weights on Hugging Face under a license with conditions. Artificial Analysis scored it 38.9, the highest result for a US lab’s open-weight model in its comparison, while several open models from Chinese labs scored higher overall. Its low hallucination rate and visual reasoning results stand out; performance on a demanding terminal-agent test was much weaker.
Muse Spark Open has 42 billion active parameters per token, with 256 experts and eight active for each token. It accepts text, images, speech and video, produces text, and has a one-million-token context window. Meta published weights in BF16 and NVFP4 formats. Its model card estimates self-hosting would require 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 a score of 38.9 on the Intelligence Index v4.3.2, rounded to 39 in its presentation. The index places it tenth among open-weight models. The nine models above it are from Chinese labs, according to the source material. Meta’s own closed Muse Spark 1.3 scored 45.1 under the same reasoning setting, while Muse Wonder scored 52.4.
The weights are available without a charge, but using the model through third-party APIs is not free. Four providers listed 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 rates match the reported list price for closed Muse Spark 1.3. Artificial Analysis calculated a weighted cost of $0.84 for one Intelligence Index task; that figure reflects its benchmark workload, not a universal cost per user request.
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 Bigger US Open Model, With Limits
The release gives developers access to a large model from Meta’s flagship line, changing the scale of Meta’s recent open-weight offerings. The source compares Muse Spark Open with Muse Glimmer, a 30-billion-parameter model that scored 17 on the current index, and Llama 4 Maverick, which scored 9. On the reported results, Muse Spark Open also exceeds NVIDIA’s same-day Nemotron 3.5 Super VL, which scored 26.5.
That US comparison does not make it the open-weight field’s leader. Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3, and other models from Chinese labs also placed above it. The result is relevant to teams weighing access, performance, cost and license terms: the model’s score alone does not settle whether it is a practical choice.
Artificial Analysis reported particular strength on its Omniscience measure, where Muse Spark Open scored 24.7 and ranked first among 167 open-weight entries. That measure rewards correct answers and penalizes confident errors. The model also recorded a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in the source comparison, and scored 81% on the visual reasoning benchmark MMMU-Pro. These results suggest possible value for research and document work, but benchmark scores do not establish performance across every real-world task.
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 Release History
Muse Spark Open is described in the source material as the first open-weight release in Meta’s Muse Spark line. It follows smaller open models, including Muse Glimmer, released in August 2026 under Apache 2.0. The material says Glimmer scored 17 on the current Intelligence Index, while Llama 4 Maverick, released in April 2025, scored 9.
The model’s license has conditions, according to the source, but the supplied material does not give its full text or detail the restrictions. That distinction matters because several competing models cited in the comparison use MIT licenses, while others have their own licenses. Developers need to review the published terms before deciding how to use or distribute the model.
The comparison also depends on a particular benchmark and test setup. Artificial Analysis tested Muse Spark Open on the day of release at xhigh reasoning effort. Its results are useful for comparing models under that setup, but they do not measure every workload or deployment environment.
““built for coding and agentic work””
— Meta, in the Muse Spark Open model card
large memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Workload Questions
The supplied source does not include the full license text, so the conditions on use and redistribution cannot be assessed from this material alone. It also does not establish how the model performs across production workloads beyond the cited tests.
Agent results vary sharply by benchmark. Artificial Analysis recorded 76% on Terminal-Bench 2.1 but 8% on Terminal-Bench 4.0, the harder version used in the Intelligence Index. The source also reports 50% on AutomationBench-AA, compared with 69% for DeepSeek V4.1 Flash. These figures show different outcomes across tests; they do not by themselves explain the causes or predict results for a particular team’s agent setup.
API rates and the weighted benchmark cost are reported figures, but actual spending will depend on provider, token volume, caching and the task. The source excerpt also ends during its discussion of input-token use, so it does not provide the rest of that comparison.
As an affiliate, we earn on qualifying purchases.
Review Terms Before Deployment
Developers can inspect the weights on Hugging Face and compare the model card’s hardware guidance with their own infrastructure. Before deploying it, they will need to review the license, confirm provider prices and test the model on their own prompts and workloads.
For teams considering coding agents, the reported gap between Terminal-Bench versions is a reason to evaluate the model on the specific tools and tasks they plan to run. Further evidence about real-world performance, deployment costs and use under the license will help clarify where Muse Spark Open fits alongside Meta’s closed model and competing open-weight releases.
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, and published its weights on Hugging Face.
How large is Muse Spark Open?
It has 816 billion total parameters, with 42 billion active per token. Its context window is one million tokens.
How did it score against other open-weight models?
Artificial Analysis scored it 38.9 on its Intelligence Index v4.3.2. The source says it ranked tenth among open-weight models and led the US lab models in the comparison; nine models from Chinese labs scored higher.
Is Muse Spark Open free to use?
The weights are available without a charge, but API access has per-token prices. The reported median list prices across four providers were $1.25 per million input tokens, $0.15 for cached input and $4.25 per million output tokens. The license has conditions, whose details are not included in the supplied source.
Is it strong at coding and agent tasks?
Results differ by test. Artificial Analysis recorded 76% on Terminal-Bench 2.1 and 8% on the harder Terminal-Bench 4.0. Those benchmark results do not establish how the model will perform on a particular coding or agent workload.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
