🔍 Read the full analysis: A Strong Open-Weight Model, A Closed-Model Price: Muse Spark Open on ThorstenMeyerAI.com
Get hardware and tech essentials 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, publishing its model weights on Hugging Face. Artificial Analysis scored it 38.9, but hosted access costs the same per token as the higher-scoring closed Muse Spark 1.3, and the model trails several open competitors on price and agent benchmarks.
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, but the model’s hosted token prices match those of Meta’s closed Muse Spark 1.3, which scored higher at 45.1 in the same test setting.
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. It has a one-million-token context window, accepts text, images, audio and video, and produces text. Meta published BF16 and NVFP4 weights on Hugging Face. Its model card estimates self-hosting requires about 700 GB of GPU memory for NVFP4 or 2 TB for BF16, and says it is optimized to run on two Mac Studios with M5 Ultra using MLX.
Artificial Analysis tested the model at its default xhigh reasoning effort. It reports an Intelligence Index score of 38.9, rounded to 39. The source describes that as the highest score from a US lab’s open-weight model and tenth among open-weight models overall; the nine models ahead are from Chinese labs. NVIDIA’s Nemotron 3.5 Super VL, tested on the same day, scored 26.5, according to Artificial Analysis.
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 closed Muse Spark 1.3. The weights can be downloaded without charge, but running the model through these providers is billed by token.
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
Open Weights, Closed-Model Pricing
The release gives developers access to Meta’s large Muse model weights, allowing them to inspect, adapt or self-host the system rather than rely only on Meta’s closed service. But access to weights is not the same as low-cost deployment: self-hosting calls for substantial GPU memory, while the listed hosted rates match the closed Muse Spark 1.3.
That distinction matters for teams weighing control against operating cost. The source reports an Artificial Analysis weighted cost of $0.84 per Intelligence Index task for Muse Spark Open. It lists MiMo-V2.6-Pro at $0.13 for a score of 46.3, and DeepSeek V4.1 Flash at $0.27 for a score of 39.5. Those comparisons are specific to the benchmark’s task mix and should not be treated as a universal estimate of application costs.
The model’s results are uneven rather than uniformly strong. Artificial Analysis reports 24% hallucination rate, the lowest among the 15 highest-scoring open-weight entries in its comparison, and 81% on the visual-reasoning benchmark MMMU-Pro, leading the 66 open-weight models tested there. But on Terminal-Bench 4.0, it scored 8%, below several named competitors. Buyers considering automated coding or shell-based agents therefore have reason to examine task-specific performance and cost rather than rely on the headline index score.
Products Worth Considering
Apple MacBook Pro 2026 14.2 Laptop with M5 Pro chip, 15‑core CPU, 16‑core GPU, 24 GB Unified Memory, 1 TB SSD, Wi‑Fi 7 – Silver
Trezor Safe 7 Crypto Hardware Wallet with Bluetooth and Color Touchscreen
Ledger Flex + Folio Crypto Hardware Wallet
As an affiliate, we earn on qualifying purchases.
Meta’s First Open Muse Model
Muse Spark Open is the first open-weight release in Meta’s flagship Muse Spark line. The source contrasts it with Meta’s earlier open releases: Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the current Artificial Analysis index, while Llama 4 Maverick, released in April 2025, scored 9. The source says Muse Spark Open is a major change in the scale of Meta’s open releases.
Its score does not put it at the top of the overall open-weight rankings. The source identifies Xiaomi’s MIT-licensed MiMo-V2.6-Pro at 46.3 as the leader, with GLM-5.3, Step 5 Preview, Kimi K3, other Qwen models and DeepSeek V4.1 Flash also scoring above Muse Spark Open. License terms vary across these models, so benchmark rankings alone do not settle which is suitable for a particular deployment.
Meta’s own closed models provide another comparison. In the same xhigh setting and at the same listed per-token prices, Muse Spark 1.3 scored 45.1, six points above the open release; Muse Wonder scored 52.4 in the figures supplied by the source. These are benchmark results, not guarantees of performance on a developer’s own workload.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card
Deployment Costs and License Terms
The supplied material does not provide the full license text or establish what commercial, redistribution or modification conditions apply. Developers should review the Hugging Face model card and license before building a product around the weights. The source characterizes the license as conditional but does not reproduce its terms.
Hosted prices are reported as a median across four third-party providers; the providers’ individual rates, service limits and pricing changes are not detailed here. Actual bills will depend on prompt length, caching, output volume and provider implementation. The cited $0.84 benchmark-task cost is tied to Artificial Analysis’s test workload, not a standard cost for every use.
The results also leave open how well the model performs in real deployments beyond the listed benchmarks. The source gives strong scores on visual reasoning and hallucination measurement but much weaker results on Terminal-Bench 4.0. It does not establish why performance differs across agent tests or whether later software, prompting or model revisions will change those results.
Check the Model Before Deployment
Developers can now examine and download the BF16 and NVFP4 weights on Hugging Face, subject to the release’s license terms. Teams evaluating hosted use can compare provider rates and test their own workloads, including cached-input behavior and long-running agent tasks, before estimating total cost.
Further clarity will depend on the full licensing details, provider-specific pricing and independent testing across additional applications. For now, the published results give potential users a starting point: a high-scoring US open-weight model with notable visual-reasoning and hallucination results, but not a clear price-performance leader against several open competitors.
Key Questions
What is Muse Spark Open?
It is Meta’s 816-billion-parameter mixture-of-experts model, released as open weights on October 15, 2026. Meta lists 42 billion active parameters per token and a one-million-token context window.
How did Muse Spark Open score?
Artificial Analysis measured 38.9 on its Intelligence Index v4.3.2 at xhigh reasoning effort; the score is rounded to 39. The supplied source places it tenth among open-weight models overall.
Is hosted access cheaper than Meta’s closed model?
Not at the reported list prices. Four third-party providers had a median of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens—the same rates cited for Muse Spark 1.3.
Can developers run it on their own hardware?
Meta’s model card estimates about 700 GB of GPU memory for NVFP4 weights and 2 TB for BF16. It also says the model is optimized for two Mac Studios with M5 Ultra using MLX; actual setup requirements may depend on the configuration.
Is it strong for coding agents?
Results differ by test. The source reports 76% on Terminal-Bench 2.1 but 8% on Terminal-Bench 4.0, a harder benchmark included in the Artificial Analysis index. Those results do not establish how it will perform on every coding or agent workload.
Source: ThorstenMeyerAI.com
