🔍 Read the full analysis: How Muse Spark Open Stacks Up On Strength And Price on ThorstenMeyerAI.com
Get the little things that make your day 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, its first open-weight model in the Muse Spark line. Artificial Analysis scored it 38.9, with leading open-model results in hallucination resistance and visual reasoning, but the benchmark’s weighted task cost was $0.84—far above several competitors. Its weights are free to download, while API pricing matches closed Muse Spark 1.3; the license has conditions.
Meta released Muse Spark Open on October 15, giving developers access to the first open-weight model in its flagship Muse Spark line. Artificial Analysis scored it 38.9 on its Intelligence Index, while its weighted cost per benchmark task was $0.84, more than several open-weight rivals with similar or higher scores.
The model has 816 billion total parameters, with 42 billion active per token, and its weights are available on Hugging Face. It accepts text, images, speech or audio, and video, and can produce text. The context window is 1 million tokens. Meta provides BF16 and NVFP4 weights; its model card estimates self-hosting requires about 2 terabytes of GPU memory for BF16 or 700 gigabytes for NVFP4.
Artificial Analysis tested Muse Spark Open at its default xhigh reasoning effort. Its 38.9 score makes it the highest-scoring open-weight model from a US lab in the source’s comparison and tenth overall. Nine open-weight models scored higher, all from Chinese labs. The comparison is specific to this version of the index and its tested models.
The weights can be downloaded without a per-token charge, but access through four third-party API providers carried median listed prices 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 listed prices for closed Muse Spark 1.3. Artificial Analysis calculates a separate $0.84 weighted cost per index task; that benchmark figure reflects the model’s token use during the evaluation, not a flat fee for every user task.
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 Research Scores, Costly Runs
The results point to a model with a particular strength: answering accurately across demanding knowledge and visual tasks. Artificial Analysis gave it 24.7 on AA-Omniscience, the highest score among the 167 open-weight entries in that test. The benchmark rewards correct answers and penalizes confident errors. Its reported hallucination rate was 24%, lowest among the 15 highest-scoring open-weight models in the comparison. On MMMU-Pro, a visual reasoning benchmark, it scored 81%, the best result among 66 tested open-weight models.
Those results may make the model relevant to document analysis and research workflows, especially where long inputs and visual information matter. They do not establish that it will be the best choice for every research task: benchmark scores cover defined test sets, and the source provides no independent measurement of performance across everyday customer workloads.
For buyers comparing hosted models, the task-cost results complicate the headline score. Artificial Analysis put MiMo-V2.6-Pro at 46.3 for $0.13 per index task, while DeepSeek V4.1 Flash scored 39.5 for $0.27. Muse Spark Open scored below the former and close to the latter at a higher benchmark cost. Its listed API rates also match Muse Spark 1.3, which scored 45.1 at the same reasoning effort in the source’s comparison. Free weights and paid API service are distinct options, and the model’s license conditions matter to anyone planning to use or redistribute it.
Performance was uneven across agent tests. Meta describes the model as built for coding and agentic work, and the source reports scores of 89.2 on MCP-Atlas, 80.4 on OSWorld Verified and 76.4% on Terminal-Bench 2.1. But on Terminal-Bench 4.0, the harder test included in the Intelligence Index, Artificial Analysis measured 8%. It also reported 50% on AutomationBench-AA and Elo scores below the cited leaders on GDPval-AA and AA-Briefcase. This split matters for developers considering long-running shell agents, where unsuccessful runs can consume tokens without completing the job.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How It Compares With Open Rivals
Muse Spark Open marks a change in scale for Meta’s open releases. Muse Glimmer, released in August 2026, has 30 billion parameters and scored 17 on the same index, according to the source. Llama 4 Maverick, released in April 2025, scored 9. The new model’s 816-billion total parameter count brings Meta into a larger class of open-weight releases, but size alone does not determine benchmark ranking or operating cost.
Within the US open-weight group, the source also compares it with NVIDIA’s Nemotron 3.5 Super VL, released on the same day, which scored 26.5. Muse Spark Open’s 38.9 is higher, though the source characterizes NVIDIA’s model as permissively licensed and Muse Spark Open’s license as conditional. The exact rights and restrictions depend on the applicable license terms; the available source does not spell them out.
The strongest open-weight scores in the source’s overall comparison came from models including Xiaomi’s MIT-licensed MiMo-V2.6-Pro at 46.3, Step 5 Preview at 43.7, Kimi K3 at 43.6 and GLM-5.3 at 44.8. Several of these have their own license terms. The cited prices and benchmark costs are the source’s reported figures, and actual bills can vary with provider, caching, workload and token volume.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card
large capacity GPU memory for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License Terms and Real-World Costs
The source describes Muse Spark Open’s license as having conditions but does not reproduce those terms. Developers should consult the Hugging Face model card and license before deployment, redistribution or commercial use. The source also says the weights are free to download; that does not by itself settle what uses the license permits.
The benchmark figures do not show how the model will perform on an individual company’s workload or what its final bill will be. The $0.84 index-task cost is based on Artificial Analysis’s evaluation, and token use, caching and provider rates can change costs. The source reports that input accounted for $0.67 of the benchmark task cost, including cache reads and writes, but its comparison does not establish a universal cost for typical prompts.
It is also unclear from the available material how the model’s strong scores on selected tests will translate to routine coding and agent deployments. The 8% Terminal-Bench 4.0 result is a warning about that test category, not proof that every shell-based task will fail. The source does not provide enough information to explain the gap between that result and the higher Terminal-Bench 2.1 score.
As an affiliate, we earn on qualifying purchases.
What Developers Should Check
Developers weighing the model can compare its performance on their own representative prompts with the specific benchmark results, while tracking input, cached input and output tokens separately. Those figures will help determine whether its long context and accuracy results justify its API expense for a given workload.
Before building on the downloadable weights, teams should review Meta’s current model card and license for the rights and conditions that apply to their intended use. API availability and provider pricing can also change, so the listed rates should be checked with the provider at the time of deployment. Further benchmark runs or updates may clarify how the model performs on agent tasks and whether the gap between the two Terminal-Bench versions persists.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is Meta’s first open-weight model in the Muse Spark line, released on October 15, 2026. It has 816 billion total parameters, 42 billion active per token, and a 1-million-token context window.
How did it score against open-weight models?
Artificial Analysis reported an Intelligence Index score of 38.9 at xhigh reasoning effort. The source ranks it tenth among open-weight models overall and first among US open-weight models in its comparison. Nine models, all from Chinese labs, scored higher overall.
Are the weights free, and what does API access cost?
The weights are available on Hugging Face without a per-token charge, subject to the model’s license conditions. The source reports median listed API prices of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens across four providers.
Is Muse Spark Open a good choice for coding agents?
The results are mixed. The source reports 76.4% on Terminal-Bench 2.1, but Artificial Analysis measured 8% on Terminal-Bench 4.0, the harder version used in the Intelligence Index. Teams should treat those as benchmark results and test the model on their own agent tasks.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
