🔍 Read the full analysis: Muse Spark Open: Strength In Open Weights, Pricing Like Meta on ThorstenMeyerAI.com
TL;DR
Meta released Muse Spark Open on October 15, 2026, with weights on Hugging Face and an Artificial Analysis Intelligence Index score of 38.9. It leads tested open-weight models on the AA-Omniscience and MMMU-Pro measures, but its token prices match the higher-scoring closed Muse Spark 1.3, and it performed poorly on Terminal-Bench 4.0. Its license has conditions, and the source material does not detail them.
Meta released Muse Spark Open, an 816-billion-parameter open-weight model, on October 15, putting its weights on Hugging Face and pricing API access at the same list rates as its closed Muse Spark 1.3. Artificial Analysis measured a score of 38.9 on its Intelligence Index, while reporting stronger results in factuality and visual reasoning than other tested open-weight models. The release gives developers access to Meta’s largest open-weight model in the Muse Spark line, but the benchmark and price comparisons leave open questions about where it makes practical sense to use.
The model uses a mixture-of-experts design with 816 billion total parameters and 42 billion active per token. Meta lists a one-million-token context window and text, image, speech and video inputs; it generates text. The weights are available in BF16 and NVFP4 formats. Meta’s 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 for two Mac Studios with M5 Ultra using MLX.
Artificial Analysis tested Muse Spark Open at its default xhigh reasoning effort on October 15. Its Intelligence Index v4.3.2 score was 38.9, rounded to 39 by the evaluator. The source report says that ranks it tenth among open-weight models overall and highest among those from a US lab. Nine models ahead of it are from Chinese labs. Those rankings describe the tested set and index, rather than every possible model or use case.
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 reported for Muse Spark 1.3. Artificial Analysis estimated a weighted cost of $0.84 per index task for Muse Spark Open. Of that, $0.67 was input-token cost: $0.40 for cache reads, $0.25 for cache writes and $0.02 for uncached input; output accounted for $0.17.
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
Strengths Come With Tradeoffs
The release expands the options for developers who want to inspect and run a large Meta model using its published weights. Its reported 24.7 AA-Omniscience score was the highest among 167 open-weight entries in the source’s comparison. Artificial Analysis describes this measure as rewarding correct answers and penalizing confident errors. The model also scored 81% on MMMU-Pro, which the report places first among 66 tested open-weight models. Its 24% hallucination rate was the lowest among the 15 highest-scoring open-weight entries in that comparison.
Those results may suit research and document work that combines long context with images, but they do not make the model a universal leader. On Terminal-Bench 4.0, Artificial Analysis measured 8%, compared with 33% for Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3. The source report also gives Muse Spark Open a lower GDPval-AA score than MiMo-V2.6-Pro and Muse Spark 1.3. These differences matter to teams choosing models for tasks that involve operating software or completing multi-step work.
Price and access also pull in different directions. The weights are available without a token charge, but using hosted APIs costs the same per token as Meta’s closed Muse Spark 1.3, which scored 45.1 on the same index at the same reasoning effort. Self-hosting requires substantial memory, while several open-weight competitors in the report had lower API costs per index task. Buyers will need to weigh control over the weights, infrastructure demands, license terms and performance on their own workloads.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Line Expands
Muse Spark Open is the first open-weight release in Meta’s flagship Muse Spark line, according to the source report. Meta’s recent open models were smaller: Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the current index; Llama 4 Maverick, released in April 2025, scored 9. The new model changes the scale of Meta’s open releases, while its overall score remains below the leading open-weight models cited in the report.
Artificial Analysis scored NVIDIA’s Nemotron 3.5 Super VL, released the same day, at 26.5. The report calls it the leading US open-weight model among permissively licensed options; Muse Spark Open scored higher, but its license has conditions. In the broader open-weight comparison, Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3. The source also lists several other models ahead of Muse Spark Open, including GLM-5.3, Step 5 Preview, Kimi K3, Qwen3.8 models and DeepSeek V4.1 Flash.
The source gives Muse Spark Open a 1-million-token context window and reports 2.96 billion input tokens consumed across the Intelligence Index run. It says that is fewer than MiMo-V2.6-Pro’s 4.43 billion. The index-task cost figures are weighted estimates from that evaluation, not a guarantee of what a particular customer will pay: actual API usage depends on the workload and provider.
„“built for coding and agentic work”“
— Meta, in its model card
large language model hosting hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Workload Questions
The source material says the license has conditions but does not spell out those conditions. Developers should review the published license before deciding whether the model fits their intended use. It also does not identify the four API providers or provide a provider-by-provider price breakdown beyond the reported median list prices.
Benchmark scores are measurements under specified tests and settings. The report identifies xhigh reasoning effort for the Intelligence Index run, but it does not establish how the model will perform across customer workloads, deployment setups or different reasoning settings. The cited token prices and weighted task cost likewise do not determine the bill for an individual application.
AI model training GPU requirements
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Check Terms and Test Workloads
The weights are available on Hugging Face, and third-party API providers are listing access at the reported rates. The next practical step for developers is to review Meta’s license and model card, compare the memory requirements with available hardware, and test the model against their own tasks. Teams considering coding or autonomous tool use have reason to pay particular attention to results on the relevant terminal and agent benchmarks.
The source material does not give a date for a further Meta release, a change to pricing or an update to the license. For now, the published weights, provider prices and evaluation results are the available basis for deciding whether Muse Spark Open fits a particular deployment.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released it on October 15, 2026, and Artificial Analysis tested it the same day.
What did Muse Spark Open score?
It scored 38.9 on Artificial Analysis Intelligence Index v4.3.2, which the evaluator rounds to 39. The source report ranks it tenth among open-weight models overall and highest among open-weight models from a US lab.
Is Muse Spark Open free to use?
Meta published the weights on Hugging Face, but hosted API use is billed by tokens. The reported median list prices were $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens.
How does it compare with Muse Spark 1.3?
At the same xhigh reasoning effort, the source reports Muse Spark 1.3 scored 45.1, compared with 38.9 for Muse Spark Open. Their reported API list prices are the same.
What remains unknown about the release?
The source material does not detail the conditions in Muse Spark Open’s license or identify the four API providers. Benchmark results also do not establish performance or cost for a specific customer workload.
Source: ThorstenMeyerAI.com
