🔍 Read the full analysis: Muse Spark Open Matches Open Weights With A Closed-Model Price Point 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, an 816-billion-parameter open-weight model, on October 15, 2026. Artificial Analysis scored it 38.9, while listed hosted rates match Meta’s higher-scoring closed Muse Spark 1.3; the open model also performed unevenly on agent tests. The weights can be downloaded, but the model’s license has conditions and self-hosting calls for substantial hardware.
Meta released Muse Spark Open, an 816-billion-parameter open-weight model, on Hugging Face on October 15, giving developers access to its weights while setting hosted API prices at the same list rates as the higher-scoring closed Muse Spark 1.3. Artificial Analysis measured a 38.9 Intelligence Index score for the open model, and reported uneven results on agent tasks, including an 8% score on Terminal-Bench 4.0.
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. Meta lists a one-million-token context window and support for text, images, speech or audio, and video as inputs; its output is text. The Hugging Face release includes BF16 and NVFP4 weights.
Artificial Analysis tested the model at its default xhigh reasoning effort and recorded 38.9 on the Intelligence Index v4.3.2, rounded to 39 in the source report. That placed it tenth among open-weight models overall and at the top among open-weight models from US labs in the report’s comparison. The nine models above it came from Chinese labs.
Meta’s model card estimates that self-hosting needs about 700 GB of GPU memory for NVFP4 weights or 2 TB for BF16. It also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX. For hosted access, four third-party providers had a median listed price of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens—the same listed rates as Muse Spark 1.3.
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 changes Meta’s open-model offering by bringing a large model from its Muse Spark flagship line into an open-weight format. Developers can download the weights and, subject to the model’s license, run them on their own systems. That can offer more control over deployment than a closed model’s API, but the stated memory requirements put self-hosting beyond many individual developers and smaller teams.
For users who rent access instead, the price comparison is less favorable. The source report says a weighted Artificial Analysis Index task cost $0.84 on Muse Spark Open. It lists MiMo-V2.6-Pro at a score of 46.3 for $0.13 per task and DeepSeek V4.1 Flash at 39.5 for $0.27. These are reported benchmark and cost figures, not a guarantee of performance or cost in a particular workload. They indicate that hosted access to Muse Spark Open is not the lowest-cost option among the compared open-weight models.
The score also needs to be read alongside the model’s strengths and limits. The report gives it a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison, and strong visual-reasoning results. But its Terminal-Bench 4.0 result was 8%, well below the cited results for Muse Spark 1.3 and several open competitors. That gap matters to teams evaluating autonomous coding or shell-based agents, where a poor run can consume both time and paid tokens.
high-performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Muse Spark Ranks
Muse Spark Open is Meta’s first open-weight release in the Muse Spark line, according to the source report. Meta’s recent open models were smaller: Muse Glimmer, released in August 2026, has 30 billion parameters and scored 17 on the same current index, while Llama 4 Maverick scored 9 in the report’s comparison. The new model’s scale is a substantial change for Meta’s open-weight releases, though it does not put the model at the top of the wider field.
Artificial Analysis scored the closed Muse Spark 1.3 at 45.1 at the same reasoning effort and list price, six points above Muse Spark Open. Muse Wonder scored 52.4 in the source report. Among open-weight models, MiMo-V2.6-Pro led the cited comparison at 46.3; other models from Chinese labs also ranked ahead of Meta’s release. These rankings depend on the named benchmark version, test settings and comparison set.
The report also describes results that could make Muse Spark Open useful for document-heavy or research work: 81% on MMMU-Pro, which it ranked first among the 66 tested open-weight models, and 78% on AA-LCR long-context reasoning. Its one-million-token context window adds capacity for long inputs, but a large context limit alone does not establish accuracy or lower costs on a given task.
“built for coding and agentic work”
— Meta, in its Muse Spark Open model card
large memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Workload Limits
The source report says the model’s weights are available on Hugging Face and describes its license as having conditions, but the supplied material does not include the license terms. Users should review the license itself before deploying or modifying the model. The report does not establish that the weights are available under a permissive license.
It is also unclear how the reported provider prices will change over time, whether they apply uniformly to every deployment, or how performance will vary across real-world tasks. The source gives a weighted cost for one benchmark index run, including its test composition; it does not provide a universal cost estimate. Several benchmark figures are reported, but no details are supplied here about independent replication, error ranges or the full testing methodology.
Agent performance is particularly uncertain beyond the cited benchmarks. The source reports 76% on Terminal-Bench 2.1 but 8% on the harder Terminal-Bench 4.0, alongside results from other agent evaluations. Those figures point to uneven performance, not a definitive answer about every coding or automation use. The source material’s cost breakdown also ends mid-sentence after beginning a comparison of input-token use, so no further cross-model token comparison can be confirmed from it.
As an affiliate, we earn on qualifying purchases.
What Developers Should Check
Developers considering the model can compare the license terms, hardware requirements and provider pricing with their intended deployment before choosing between self-hosting and an API. For an API evaluation, the relevant cost will depend on the volume and type of tokens used, including cached input, as well as the provider’s current rates.
Teams focused on research, visual reasoning or long documents can test those tasks directly against their own data and compare results with other models. Teams planning shell-based agents should pay particular attention to the Terminal-Bench 4.0 result and run controlled tests before assigning unattended work. The source material does not identify a further Meta announcement or a scheduled benchmark update, so the next concrete developments are likely to be additional provider availability, independent evaluations and practical testing by users.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta release?
Meta released Muse Spark Open on Hugging Face on October 15, 2026. It is an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token.
How did Muse Spark Open score?
Artificial Analysis measured a 38.9 score on its Intelligence Index v4.3.2, tested at default xhigh reasoning effort. The source report says that rounds to 39.
Is the model free to use?
The weights are available for download, but the source describes the license as having conditions and does not provide its full terms. Hosted access is not free: 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 much hardware does self-hosting require?
Meta’s model card estimates about 700 GB of GPU memory for NVFP4 weights or 2 TB for BF16. It also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX.
Is Muse Spark Open a strong choice for coding agents?
The reported results are mixed. The model scored 76% on Terminal-Bench 2.1 but 8% on the harder Terminal-Bench 4.0, so the cited results do not establish reliable performance across agent tasks. Developers should test the model on their own workflows before relying on it for unattended runs.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
