🔍 Read the full analysis: Muse Spark Open: Open Weights, Strong Performance, Meta-Like Pricing on ThorstenMeyerAI.com
Get smart everyday buys 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, making its 816-billion-parameter model available as downloadable weights. Artificial Analysis measured a 38.9 Intelligence Index score and strong results on visual reasoning and its hallucination test, but weaker performance on a demanding agent benchmark. Hosted API list prices match Meta’s closed Muse Spark 1.3, which scored higher in the same evaluation.
Meta released Muse Spark Open, an 816-billion-parameter mixture-of-experts model, on Hugging Face on October 15, 2026, giving developers access to its weights. Artificial Analysis recorded a 38.9 score on its Intelligence Index, while listed prices from four third-party providers matched those of Meta’s closed Muse Spark 1.3, which scored higher in the same test.
The model has 42 billion active parameters per token, uses 256 experts with eight active for each token, and supports a one-million-token context window. Its input modalities include text, images, speech and video; it generates text. Meta published BF16 and NVFP4 weights. Its model card estimates self-hosting requires about 700 GB of GPU memory for NVFP4 or 2 TB for BF16.
Artificial Analysis tested Muse Spark Open at its default xhigh reasoning effort. Its Intelligence Index v4.3.2 score is 38.9, rounded by the evaluator to 39. The source report places it tenth among open-weight models overall and says the nine models ahead are from Chinese labs. It also reports the model as the highest-scoring open-weight release from a US lab in that comparison; these are benchmark rankings, not a guarantee of performance on a particular workload.
The evaluation shows uneven results across tasks. Muse Spark Open scored 24.7 on AA-Omniscience, ranked first among 167 open-weight entries in that test, and had a reported 24% hallucination rate among the 15 highest-scoring open models. It also scored 81% on MMMU-Pro, a visual-reasoning benchmark. On Terminal-Bench 4.0, however, Artificial Analysis measured 8%; the source report says that was below several named open and closed competitors. Benchmark outcomes can vary with test design and settings.
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 Vision, Uneven Agent Results
The release gives developers a way to download and run a model from Meta’s flagship Muse Spark line, rather than relying only on a hosted service. That matters for teams evaluating self-hosting, data control or custom deployment, though the stated memory requirements are substantial and the source does not establish what hardware or operating costs a particular deployment would require.
The benchmark profile suggests a possible fit for visual reasoning and research or document tasks: the model combines its reported MMMU-Pro result with a long context window and a low hallucination rate in Artificial Analysis’s test. Those scores should be treated as measures of particular evaluations, not proof that it will be more reliable in every real-world use.
The results are less encouraging for autonomous terminal work. The 8% Terminal-Bench 4.0 score contrasts with higher results on other agent-related tests cited in the source, including 76% on Terminal-Bench 2.1. That gap makes the exact benchmark and task important when selecting a model. The reported cost per Intelligence Index task was $0.84, with 80% attributed to input tokens; long or unsuccessful agent runs could add expense, but actual bills depend on usage and provider rates.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Scores Compare
Muse Spark Open is a marked change in scale from Meta’s recent open releases, according to the source report. Muse Glimmer, released in August 2026, has 30 billion parameters and an Apache 2.0 license; the report gives it a current Intelligence Index score of 17. Llama 4 Maverick, released in April 2025, scored 9 in the same account. The new model therefore expands Meta’s open-weight offering, although the cited figures do not make it the overall open-model leader.
Artificial Analysis scored Meta’s closed Muse Spark 1.3 at 45.1 at the same reasoning effort, compared with 38.9 for Muse Spark Open. The source says the closed model’s listed API rates are the same as the open model’s rates from four third-party providers: a median of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. The weights themselves are free to download; hosted inference is billed separately.
Price and score comparisons in the source also vary across open models. Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3, with a reported $0.13 cost per Intelligence Index task; DeepSeek V4.1 Flash scored 39.5 at $0.27. Those figures are the source’s reported evaluation costs, not universal provider quotes. Muse Spark Open’s own license has conditions, and the source does not reproduce enough of its terms to determine what uses are permitted.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card
large capacity GPU server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Deployment Questions
The source says Muse Spark Open is subject to a license with conditions, but does not provide the full license text or specify which uses face restrictions. Developers should check the published terms directly before building a product around the weights. The source also reports Meta’s memory estimates but does not give enough detail to verify deployment costs, throughput or performance across different hardware setups.
Benchmark rankings are snapshots from specific tests, settings and providers. The source describes the Artificial Analysis evaluation as using xhigh reasoning effort, but does not provide full test logs or enough information to reproduce every result. It also gives no detailed explanation for the gap between 76% on Terminal-Bench 2.1 and 8% on Terminal-Bench 4.0. API prices are reported as a median from four providers and may change; the source does not identify those providers in the supplied material.
As an affiliate, we earn on qualifying purchases.
What Developers Should Verify
The immediate next steps are practical: review Meta’s model card and license, inspect the BF16 and NVFP4 weights on Hugging Face, and confirm whether available hardware can meet the stated memory needs. Teams considering hosted use should check current provider pricing and token accounting rather than assume the reported median remains available.
For evaluation, developers can test the model against their own workloads, especially if they are considering coding agents or long-running terminal tasks. The next useful evidence will be independent results that clarify performance across real deployments, the costs of self-hosting and the limits of the license. No additional product or benchmark milestone is specified in the source material.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta release?
Meta released Muse Spark Open on October 15, 2026. It is an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token, and its weights are available on Hugging Face.
How did Muse Spark Open score?
Artificial Analysis measured 38.9 on its Intelligence Index v4.3.2, using its default xhigh reasoning effort. The source also reports a 24.7 score on AA-Omniscience and 81% on MMMU-Pro; results depend on the test and settings.
Are the weights free, and is hosted use free?
The source says the weights are free to download, subject to the model’s license conditions. Hosted inference costs money: the reported median list prices from four providers were $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens.
Does Muse Spark Open match Meta’s closed model?
No, not on the cited Intelligence Index comparison. The source reports 38.9 for Muse Spark Open and 45.1 for closed Muse Spark 1.3 at the same reasoning effort. It says their listed API prices are the same.
What is still unknown?
The supplied source does not spell out the full license restrictions, identify the four API providers, or establish deployment costs and performance across hardware. It also does not explain the large difference between the model’s cited Terminal-Bench 2.1 and 4.0 results.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
