🔍 Read the full analysis: Muse Spark Open: An Open-Weight Model With A Meta-Like Price Tag on ThorstenMeyerAI.com
Get the latest gadgets 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, with downloadable weights for an 816-billion-parameter model. Artificial Analysis scored it 38.9, but its measured provider prices match Meta’s closed Muse Spark 1.3, and several open models score as well or better for less.
Meta released Muse Spark Open on October 15, making the first open-weight model in its flagship Muse Spark line available on Hugging Face. Artificial Analysis measured a score of 38.9 on its Intelligence Index and a median provider price of $1.25 per million input tokens and $4.25 per million output tokens—the same list prices reported for Meta’s closed Muse Spark 1.3.
The model has 816 billion total parameters, with 42 billion active for each token. Its mixture-of-experts architecture uses 256 experts, activating eight at a time. Meta lists a one-million-token context window and support for text, images, speech and video as input; output is text. The weights are available in BF16 and NVFP4 formats on Hugging Face.
Meta’s model card estimates self-hosting requires 700 GB of GPU memory for NVFP4 or 2 TB for BF16. It also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX. The weights can be downloaded, but teams that use third-party hosted services pay per token: Artificial Analysis reported a median list price across four providers of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens.
Artificial Analysis tested the model at its default xhigh reasoning effort. It recorded 38.9 on Intelligence Index v4.3.2, rounded to 39. The organization ranked it tenth among open-weight models overall and the highest-scoring open-weight model from a US lab in its comparison. Its ranking does not make it the field leader: the nine open-weight models above it were from Chinese labs, according to the source’s account of the index.
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 Access
The release gives developers access to weights for a model that is far larger than Meta’s recent open releases, but downloading those weights and renting the model are different cost choices. Self-hosting carries a substantial memory requirement; hosted use comes at prices matching Meta’s closed Muse Spark 1.3. That makes the word “open” relevant to access and deployment options, while the price comparison remains important for teams choosing a service.
Performance also varies by task. Artificial Analysis gave Muse Spark Open a score of 24.7 on AA-Omniscience, the top result among 167 open-weight entries in that test, which rewards correct answers and penalizes confident wrong ones. It measured a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison. On visual reasoning, its 81% result on MMMU-Pro led 66 tested open-weight entries.
Those results make the model relevant to document-heavy and research work, especially alongside its long context window. They do not establish that it is the best choice for every workload. For example, Artificial Analysis measured just 8% on Terminal-Bench 4.0, a test of terminal-based agent tasks. The mixed results mean buyers need to compare performance and cost on the tasks they actually plan to run.
As an affiliate, we earn on qualifying purchases.
A New Scale for Meta’s Open Models
Muse Spark Open is a substantial change in scale for Meta’s recent open-weight releases. The source compares it with Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0 that scored 17 on the current index, and Llama 4 Maverick, released in April 2025, which scored 9. NVIDIA’s Nemotron 3.5 Super VL, released on the same day as Muse Spark Open, scored 26.5 in the comparison.
The new model’s score also falls short of Meta’s closed offerings cited in the source. At the same xhigh effort and list price, Muse Spark 1.3 scored 45.1, while Muse Wonder scored 52.4. In the wider open-weight field, Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3. The source also lists DeepSeek V4.1 Flash at 39.5, close to Muse Spark Open’s 38.9.
Artificial Analysis calculated a weighted cost of $0.84 per Intelligence Index task for Muse Spark Open. The source attributes $0.67 of that amount—80%—to input tokens: $0.40 for cache reads, $0.25 for cache writes and $0.02 for uncached input. Output accounted for $0.17. These figures describe the index task and its token use; they are not a universal estimate for every application.
““built for coding and agentic work””
— Meta’s model card
high memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Workload Questions
The source material does not specify the full terms of Muse Spark Open’s license, including what conditions apply to commercial use or redistribution. It describes the license as having conditions and contrasts it with permissively licensed models, but readers should consult the model’s license before deciding whether it fits a particular deployment.
Benchmark scores do not establish how the model will perform on a team’s own prompts, data or production workload. The available figures also do not show whether hosted prices will change, or how performance and cost compare across different hardware and serving setups. Those details remain unsettled in the source material.
As an affiliate, we earn on qualifying purchases.
Check Terms and Task Results
Developers can review the weights and model card on Hugging Face, check the license terms, and compare provider prices before committing to a deployment. Teams evaluating it for research or document work can test those tasks directly, while teams considering autonomous terminal work should pay particular attention to the gap between the reported agent benchmarks.
Further provider pricing and independent evaluations could clarify how the model performs outside the Intelligence Index. Until then, the confirmed picture is a large downloadable model with strong results on some tests, weaker results on others, and hosted prices matching Meta’s closed Muse Spark 1.3.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released it on October 15, 2026. Its weights are available on Hugging Face.
How did Muse Spark Open score?
Artificial Analysis measured 38.9 on Intelligence Index v4.3.2 at its default xhigh reasoning effort. The index rounds that result to 39.
Is hosted use free because the weights are available?
No. The weights are downloadable, but the source reports a median third-party list price of $1.25 per million input tokens and $4.25 per million output tokens. Cached input was listed at $0.15 per million tokens.
What hardware does self-hosting require?
Meta’s model card estimates 700 GB of GPU memory for NVFP4 or 2 TB for BF16. It also describes an optimized setup using two Mac Studios with M5 Ultra and MLX.
Is Muse Spark Open a strong choice for terminal agents?
The reported results are mixed. Artificial Analysis measured 8% on Terminal-Bench 4.0; the source also reports 76% on Terminal-Bench 2.1. Those figures concern different test versions and do not establish performance on every agent workload.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
