🔍 Read the full analysis: How Muse Spark Open Pairs Open Weights With Meta’s Closed-Model Pricing on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Meta released Muse Spark Open on October 15, 2026, publishing its weights on Hugging Face under a license with conditions. Artificial Analysis scored it 38.9, but Meta’s closed Muse Spark 1.3 scored 45.1 at the same list price; hosting through providers costs a median $1.25 per million input tokens and $4.25 per million output tokens.
Meta released Muse Spark Open on October 15, publishing the weights for an 816-billion-parameter model on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, while Meta’s closed Muse Spark 1.3 scored 45.1 at the same list price—making the release significant both for access to Meta’s flagship model family and for the costs of using it through providers.
Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token and 256 experts, eight active for each token. It accepts text, images, speech and video, produces text, and has a one-million-token context window. Meta published BF16 and NVFP4 weight formats. 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. Its 38.9 score, rounded to 39, is the highest among open-weight models from US labs in the index, according to the source report. It ranks tenth among open-weight models overall; the nine ahead are attributed to Chinese labs. The index score is a benchmark result, not a measure of performance on every task or in every deployment.
Four third-party 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 are the same list prices as closed Muse Spark 1.3. The model weights are available to download, but access through those providers is paid. Meta’s license also has conditions, so “open weights” does not mean unrestricted use.
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
The Cost of Open-Weight Access
The release gives developers access to a large Meta model’s weights, allowing them to evaluate self-hosting rather than relying solely on a hosted API. That can matter for teams with specific infrastructure, data-handling or deployment needs. The hardware estimates, however, put self-hosting beyond the reach of many individual developers and smaller organizations.
For users who rent access, the price comparison is less favorable: the listed per-token rates match closed Muse Spark 1.3, which scored higher on the same index. The source report also gives a weighted cost of $0.84 per Intelligence Index task for Muse Spark Open. It lists MiMo-V2.6-Pro at $0.13 per task with a score of 46.3, and DeepSeek V4.1 Flash at $0.27 with a score of 39.5. Those figures suggest buyers should compare particular workloads and providers, not treat downloadable weights as proof that inference will be cheap.
Performance varies by task. Artificial Analysis reported strong results for visual reasoning and resistance to hallucination: 81% on MMMU-Pro and an AA-Omniscience score of 24.7, both described in the source as leading among the tested open-weight models. But the model scored 8% on Terminal-Bench 4.0, compared with 33% for Muse Spark 1.3. This gap cautions against assuming that a high overall score means the model is equally suitable for coding agents, research or document analysis.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Lineup
Muse Spark Open is described as the first open-weight model in Meta’s Muse Spark line. It is much larger than the company’s recent open releases cited in the source: Muse Glimmer, a 30-billion-parameter model released in August 2026, scored 17 on the current index; Llama 4 Maverick, released in April 2025, scored 9. These comparisons show how much Meta’s new model changes the scale of its open-weight offering, without establishing that it leads the full field.
The broader open-weight rankings remain led by models from Chinese developers. The source report lists MiMo-V2.6-Pro at 46.3, above Muse Spark Open’s 38.9, and says nine open-weight models overall rank ahead of Meta’s release. It also compares Muse Spark Open with NVIDIA’s Nemotron 3.5 Super VL, which scored 26.5 in the same index. Scores depend on the benchmark, test configuration and model version; they should not be read as universal rankings of quality.
Meta’s closed Muse Spark 1.3 is a direct comparison because, according to the report, Artificial Analysis tested it at the same xhigh effort and its list price matches Muse Spark Open’s provider rates. The closed model scored 45.1 versus 38.9. Muse Wonder scored 52.4 in the cited comparison, but the source does not provide enough detail here to treat that result as a like-for-like price and configuration comparison.
“The weights are free. Rented by the token, you pay closed-model rates.”
— Thorsten Meyer, ThorstenMeyerAI.com
As an affiliate, we earn on qualifying purchases.
License and Deployment Questions
The source describes the license as conditional but does not reproduce its full terms. Prospective users should check Meta’s actual license for permitted uses and restrictions; the presence of downloadable weights alone does not establish that the model is permissively licensed. The source also does not specify whether every provider offers identical settings, performance or service terms.
Benchmark scores and prices are snapshots. Artificial Analysis tested the model on release day, and provider prices can change. The source report’s discussion of input-token costs is incomplete: it says $0.67 of the $0.84 weighted cost per index task was input, including $0.40 in cache reads, $0.25 in cache writes and $0.02 in uncached input, but the supplied material ends before completing the comparison of total input-token use across models. That comparison cannot be verified from the available text.
It is also unclear how the reported benchmark results will translate to particular workloads, hardware setups or software stacks. The model card’s hardware estimates are Meta’s figures, not an independent guarantee of deployment cost or performance. No independent evaluation beyond the cited Artificial Analysis tests is included in the source material.
As an affiliate, we earn on qualifying purchases.
Testing Costs and Real Workloads
The next useful test for developers is to compare the model with alternatives on their own prompts and deployment conditions. That means checking not only benchmark scores but also quality, latency, token use and total inference cost, along with whether the model’s license fits the intended use. For teams considering self-hosting, the stated memory requirements and compatibility with their hardware will shape whether downloading the weights is practical.
Provider prices and model evaluations may change after launch, so both should be checked again before teams make a purchasing or deployment decision. The key unresolved issue is whether Muse Spark Open’s strengths in visual reasoning and hallucination measures can offset its weaker results on some agentic and coding tests at the same listed per-token rates as Muse Spark 1.3.
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.
Is Muse Spark Open free to use?
The weights are downloadable, but the source says the license has conditions. Access through third-party providers is paid, with a reported median price of $1.25 per million input tokens and $4.25 per million output tokens. Users should review Meta’s license before use.
How did it score compared with Meta’s closed model?
Artificial Analysis scored Muse Spark Open 38.9 at xhigh reasoning effort. The source reports a 45.1 score for closed Muse Spark 1.3 at the same effort and says their list prices are the same.
What tasks appear strongest or weakest in the reported tests?
The report highlights an 81% MMMU-Pro score and a 24.7 AA-Omniscience score. It also reports 8% on Terminal-Bench 4.0, indicating that the model’s results vary substantially by task and that the overall index score does not establish suitability for every agentic workload.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
