🔍 Read the full analysis: Muse Spark Open Makes A Case For Open Weights At Meta’s Price on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Spark Open on October 15, 2026, placing the 816-billion-parameter model’s weights on Hugging Face. Artificial Analysis measured a score of 38.9 and a weighted cost of $0.84 per benchmark task; the hosted token rates match Meta’s higher-scoring closed Muse Spark 1.3. Its results are strong in factuality and visual reasoning, but weak on a harder terminal-agent benchmark, and the model’s license has conditions.
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 its listed hosted token prices match those of Meta’s closed Muse Spark 1.3, which scored higher in the same comparison.
Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token and 256 experts, eight of which are active for each token. Its published inputs include text, images, audio and video; its output is text. Meta provides weights in BF16 and NVFP4, and the model card lists a one-million-token context window.
Artificial Analysis tested the model at its default xhigh reasoning effort. It recorded 38.9 on the Intelligence Index version 4.3.2, rounded to 39 in the source report. The report ranks it tenth among open-weight models overall and describes it as the highest-scoring open-weight model from a US lab. Those rankings reflect the tested model set and benchmark, not a general measure of performance in every use.
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. The source report says these rates match Muse Spark 1.3’s list prices. Artificial Analysis put the cost of one weighted Intelligence Index task at $0.84; the report attributes about 80% of that total to input tokens. The weights are available to download, but running them locally requires substantial hardware: Meta’s model card estimates 700 GB of GPU memory for NVFP4 and 2 TB for BF16.
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 Factuality, Costly Hosted Tests
The release adds a large Meta model to the open-weight field, giving developers access to downloadable weights rather than access only through a closed service. But “open-weight” does not mean unrestricted: the source report describes the license as carrying conditions, and readers should review the license itself before using the model commercially or redistributing it.
The benchmark results suggest a particular strength in factuality and visual reasoning. Artificial Analysis recorded an AA-Omniscience score of 24.7, the highest among the 167 open-weight entries in its test, and a 24% hallucination rate. It also reported 81% on MMMU-Pro, ranking first among the 66 open-weight models tested on that benchmark. These are benchmark findings, not guarantees that the model will avoid errors in production.
The economics are less favorable in the source report’s comparison with several open-weight rivals. MiMo-V2.6-Pro scored 46.3 at a reported $0.13 per index task, while DeepSeek V4.1 Flash scored 39.5 at $0.27. Muse Spark Open’s $0.84 task cost is higher in that comparison, though benchmark cost estimates depend on the test setup and provider pricing. For teams choosing a model, the central question is not just its score: it is whether its strengths justify the hosting, licensing and inference costs for their workload.
Performance also varies sharply by task. Artificial Analysis measured 8% on Terminal-Bench 4.0, a test of terminal-based agent work. The source report says that result is far below Muse Spark 1.3 and several open competitors. That gap makes broad claims about coding or agent suitability difficult to assess from headline scores alone.
Products Worth Considering
Apple MacBook Pro 14.2-Inch Laptop with M2 Pro, 16GB Memory, 1TB SSD, 19-Core GPU, Space Gray
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Midnight
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Citrus
As an affiliate, we earn on qualifying purchases.
Meta’s First Open Muse Spark
Muse Spark Open is the first open-weight release from Meta’s Muse Spark model line. The source report contrasts it with Meta’s smaller recent open models: Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the same index, while Llama 4 Maverick scored 9 in the report’s cited comparison. Benchmark versions and testing conditions matter when comparing results across releases.
The model’s overall position is mixed. According to the source report, nine open-weight models scored above it, all from Chinese labs. Meta’s closed Muse Spark 1.3 scored 45.1 at the same xhigh effort and listed token prices, while Muse Wonder scored 52.4 in the report. Those comparisons show that publishing a model’s weights does not necessarily make it the strongest version in its own product family.
There is also a distinction between an open-weight model and an open-source license. The weights can be downloaded, but the report says Muse Spark Open’s license has conditions; it does not provide enough detail here to characterize every permitted or restricted use. Developers need to consult Meta’s model card and license rather than infer permissions from the release label.
License and Real-World Fit
The source material does not spell out the license’s full conditions, so the scope of permitted commercial use, modification and redistribution cannot be established from the reported details alone. Developers should check Meta’s published license and model card before deployment.
Benchmark rankings do not establish how Muse Spark Open will perform across different prompts, systems or production workloads. The reported provider prices are median list prices from four providers, not a guarantee of what every user will pay. The source material also does not establish whether those rates or the model’s benchmark results will change after release.
Finally, the report’s benchmark comparisons are snapshots from Artificial Analysis testing on October 15. The supplied material does not include independent replication details or enough information to resolve how much results may vary with implementation, prompt design and hardware. Real-world reliability and total operating costs remain uncertain.
Check Terms Before Deployment
Developers considering the release can download the BF16 or NVFP4 weights from Hugging Face, review Meta’s model card and license, and test the model against their own tasks before committing to a deployment. Local use requires the substantial memory capacity Meta lists, while hosted use brings per-token charges at the reported rates.
Further comparisons will depend on additional testing, provider pricing and the terms attached to the weights. For teams considering agent workflows, the reported Terminal-Bench 4.0 result is a reason to run task-specific evaluations rather than rely on Meta’s broad description of the model. No later benchmark update or change to the license or pricing is established in the source material.
Key Questions
What did Meta release?
Meta released Muse Spark Open on October 15, 2026, and published its 816-billion-parameter mixture-of-experts weights on Hugging Face. It has 42 billion active parameters per token.
How did Muse Spark Open score?
Artificial Analysis measured 38.9 on its Intelligence Index v4.3.2 at xhigh reasoning effort. The source report says it ranked tenth among open-weight models in that test; rankings can change as models and evaluations are updated.
Are the weights free to use for any purpose?
The weights are downloadable, but the source report says the license has conditions. It does not provide the full terms, so users should read Meta’s license and model card before commercial use, modification or redistribution.
How much does hosted use cost?
The source report gives a median listed rate across four providers of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. It estimates $0.84 for one weighted Intelligence Index task; actual costs depend on provider rates and usage.
Is Muse Spark Open strong at agentic coding?
Results vary by test. The source report cites 76% on Terminal-Bench 2.1 but only 8% on Terminal-Bench 4.0, the harder version used in the Intelligence Index. Those results do not establish how it will perform on a particular coding or agent workflow.
Source: ThorstenMeyerAI.com
