🔍 Read the full analysis: What Makes Muse Spark Open Stand Out Among Open-Weight AI Models? 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, making its 816-billion-parameter model available as downloadable weights on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, with standout results in its hallucination and visual-reasoning tests, while reporting weaker performance on a harder terminal-agent benchmark and a weighted index-task cost of $0.84. The weights are free to download, but hosted use is priced at rates matching Meta’s closed Muse Spark 1.3.
Meta released Muse Spark Open on October 15, putting an 816-billion-parameter mixture-of-experts model on Hugging Face as downloadable weights. In same-day testing, Artificial Analysis gave it a 38.9 score on its Intelligence Index and reported leading results among open-weight models on its Omniscience and visual-reasoning measures, but weaker performance on a demanding terminal-agent test. The launch gives developers access to Meta’s largest open-weight model in the Muse Spark family, while leaving practical questions about its license, hosting requirements and cost.
Muse Spark Open has 816 billion total parameters, with 42 billion active per token and eight of 256 experts active at a time. Its model card lists a one-million-token context window and input support for text, images, speech or audio, and video; it generates text. The weights are available in BF16 and NVFP4 formats on Hugging Face. Meta estimates self-hosting requires 700 GB of GPU memory for NVFP4 or 2 TB for BF16, and says the model is optimized to run on two Mac Studios with M5 Ultra using MLX.
Artificial Analysis tested the model at its default “xhigh” reasoning effort and reported 38.9 on Intelligence Index v4.3.2, rounded to 39 in the source report. The report ranks it tenth among open-weight models overall and says the nine models ahead are from Chinese labs. Its score is below closed Muse Spark 1.3’s 45.1 under the same test setting, while the report places Muse Wonder at 52.4. These are benchmark results, not guarantees of performance on a particular task.
The evaluation gives Muse Spark Open a mixed profile. Artificial Analysis recorded 24.7 on AA-Omniscience, the highest result among 167 open-weight entries, and a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in the report. It also scored 81% on MMMU-Pro, a visual-reasoning test, ranking first among 66 open-weight entries tested. For coding and agent use, however, the same report recorded 8% on Terminal-Bench 4.0, compared with 33% for Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3.
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
Strong Research Scores, Costly Agent Runs
The release matters because it brings a large Meta model into the open-weight field, where developers can download model weights rather than rely only on a company-operated service. That can give organizations more control over deployment and data handling, subject to the license and the hardware needed. The model’s long context and reported visual and hallucination results may make it worth testing for document-heavy or research workflows, but benchmark rankings alone do not establish reliability in production.
Access is not the same as low-cost operation. The source report says 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. It says those list prices match closed Muse Spark 1.3. Artificial Analysis estimated a weighted cost of $0.84 for one Intelligence Index task; $0.67 of that was input-token charges. Those figures describe the evaluation and listed provider prices, not a universal bill for every use.
The comparison with other open-weight models also depends on what developers value. The report lists MiMo-V2.6-Pro at 46.3 on the index for $0.13 per task, and DeepSeek V4.1 Flash at 39.5 for $0.27. Muse Spark Open’s reported 38.9 score is therefore not the strongest score or lowest cost across the field. Its appeal rests on a combination of Meta’s model offering, its reported test profile and the option to self-host—not on leading every benchmark or price measure.
As an affiliate, we earn on qualifying purchases.
Meta’s First Open Muse Spark
Muse Spark Open is described in the source report as the first open-weight release from Meta’s flagship Muse Spark line. Earlier Meta open models were smaller: the report gives Muse Glimmer as a 30-billion-parameter model released in August 2026, with a score of 17 on the current index, and Llama 4 Maverick as scoring 9. On that basis, Muse Spark Open marks a substantial increase in the scale of Meta’s open-weight releases.
Its position is less clear-cut when compared with the wider field. The report says Muse Spark Open is the highest-scoring open-weight model from a U.S. lab, while placing nine models from Chinese labs above it overall. It also compares the model with NVIDIA’s Nemotron 3.5 Super VL, which the report says scored 26.5 and was released the same day. Rankings depend on the test suite and release set described by Artificial Analysis; they do not establish a model’s performance for every user or application.
There is also a meaningful distinction between open weights and an open license. The weights can be downloaded, but the source describes the license as carrying conditions and contrasts it with MIT-licensed competitors. The provided material does not set out those terms in detail, so prospective users need to review Meta’s model license before building or distributing a product around the model.
large language model hosting hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License, Hosting and Real-World Results
The source material does not provide the full license language, so it is not possible from the supplied information to specify which commercial, redistribution or deployment uses are permitted. The weights are downloadable, but organizations should not assume that availability means unrestricted use.
Self-hosting figures are Meta’s estimates, not independent hardware tests in the source. The report says the model is optimized for two Mac Studios with M5 Ultra and MLX, but gives no performance, throughput or reliability measurements for that setup. It also does not establish whether all listed providers offer the same model configuration, service terms or prices over time.
Benchmark results are snapshots, and the source does not explain how closely the tested settings match typical workloads. Muse Spark Open’s 8% Terminal-Bench 4.0 result sits alongside higher scores on other agent tests, making its suitability dependent on the task. The supplied report also ends during its discussion of input-token usage, so its full comparison of token consumption across models is unavailable.
As an affiliate, we earn on qualifying purchases.
What Developers Should Verify
Developers considering Muse Spark Open can download the weights and review Meta’s model card and license before testing it. The practical next steps are to confirm that the intended use is allowed, check whether available hardware can run the selected format, and measure performance on the team’s own prompts and evaluation set.
Teams assessing hosted access should check current provider prices, including treatment of cached input and output tokens, because the reported figures are list-price snapshots. For agent workflows, testing should include the specific tools and terminal tasks a system will face, alongside controls for failed or repeated runs. Further independent evaluations and updated provider and hardware information would help clarify how the model performs beyond the initial benchmark results.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is Meta’s 816-billion-parameter open-weight model, released on Hugging Face on October 15, 2026. It uses a mixture-of-experts design with 42 billion parameters active per token.
How did it score on Artificial Analysis’s Intelligence Index?
Artificial Analysis reported a score of 38.9 on Intelligence Index v4.3.2 at its default “xhigh” reasoning effort. The source report rounds this to 39; it ranks the model tenth among open-weight models overall.
Is Muse Spark Open free to use?
The weights are available to download, but the source describes the license as conditional and does not give its full terms. Hosted use is not free in the listed examples: the report gives median prices of $1.25 per million input tokens and $4.25 per million output tokens across four providers. Review the license and provider terms before use.
Is it suitable for coding agents?
The model card describes it as built for coding and agentic work, and the report records some strong agent-related results. But its 8% score on Terminal-Bench 4.0 was substantially below several comparison models. Teams should test their own agent tasks rather than treat any single benchmark as decisive.
What hardware does self-hosting require?
Meta’s model card estimates 700 GB of GPU memory for NVFP4 and 2 TB for BF16. Meta also says the model is optimized to run on two Mac Studios with M5 Ultra using MLX; the supplied material does not provide independent performance results for that setup.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
