🔍 Read the full analysis: What Sets Muse Spark Open Apart On Model Strength And Cost? 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 weights available on Hugging Face. Artificial Analysis scored it 38.9, citing strong results in visual reasoning and factual reliability, but its per-task inference cost was $0.84—far above some open-weight competitors with similar or higher scores. Its license has conditions, and performance on a demanding agent benchmark was weak.
Meta released Muse Spark Open, an 816-billion-parameter open-weight model, on October 15, 2026, and Artificial Analysis scored it 38.9 out of 100 on its Intelligence Index. The model performs particularly well on visual reasoning and a test designed to measure factual reliability, but its measured task cost was $0.84—well above some open-weight rivals that score similarly or higher. The release gives developers access to Meta’s flagship Muse model weights, though its license has conditions and its capabilities vary sharply by task.
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. Meta’s model card lists a one-million-token context window and support for text, images, speech or audio, and video as inputs, with text output. BF16 and NVFP4 weights are available on Hugging Face. Meta estimates self-hosting requires about 700 GB of GPU memory for NVFP4 weights or 2 TB for BF16.
Artificial Analysis tested the model at its default xhigh reasoning effort. Its Intelligence Index score of 38.9 is rounded to 39 in the report. The source describes it as the highest-scoring open-weight model from a US lab and tenth overall among open-weight models tested; the nine models ahead of it were from Chinese labs. The score is below Meta’s closed Muse Spark 1.3, which received 45.1 under the same effort setting, according to the supplied comparison.
The reported strengths include an AA-Omniscience score of 24.7, first among 167 open-weight entries, and a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in the comparison. It also scored 81% on MMMU-Pro, a visual-reasoning benchmark, placing first among 66 open-weight models tested. These are benchmark results, not a guarantee of performance in every real-world use.
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
Strengths Come With a Cost
The release shifts the scale of Meta’s open-weight offerings: an 816-billion-parameter model is a much larger entry than the company’s recent open releases, including 30-billion-parameter Muse Glimmer. For teams handling long documents, research material, or visual inputs, its one-million-token context window and benchmark results may make it worth testing.
But open weights do not mean low-cost inference. The source reports a median provider list 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 closed Muse Spark 1.3. In Artificial Analysis’s index run, Muse Spark Open cost $0.84 per task. Its reported cost was $0.67 for input, including cache reads and writes, and $0.17 for output.
That price matters because some competitors delivered comparable index scores for less. The supplied figures put MiMo-V2.6-Pro at 46.3 for $0.13 per index task, and DeepSeek V4.1 Flash at 39.5 for $0.27. Those comparisons reflect a specific benchmark and pricing snapshot, not a universal cost for every workload. Still, they show why teams should weigh task performance, license terms, and usage costs together rather than treating the weights’ availability as the whole cost picture.
high performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How It Ranks Against Rivals
Muse Spark Open is the first open-weight release in Meta’s flagship Muse Spark line. It is a notable change in scale from Muse Glimmer, released in August 2026, which the source says scored 17 on the current Artificial Analysis index. Llama 4 Maverick scored 9 in the comparison. NVIDIA’s Nemotron 3.5 Super VL, released on the same day as Muse Spark Open, scored 26.5, according to the source.
The broader open-weight leaderboard remains led by models from Chinese labs. The cited top score is 46.3 for Xiaomi’s MIT-licensed MiMo-V2.6-Pro; the supplied comparison also lists GLM-5.3, Step 5 Preview, Kimi K3, and other models above Muse Spark Open. The 38.9 result therefore makes Meta’s model a leading US open-weight entry in this comparison, but does not put it at the top of the overall field.
Results also depend on the task. The source reports 76% on Terminal-Bench 2.1, but only 8% on Terminal-Bench 4.0, the harder version used in the Intelligence Index. It lists Muse Spark 1.3 at 33% on Terminal-Bench 4.0, with MiMo-V2.6-Pro at 35% and GLM-5.3 at 42%. The figures illustrate that a high overall score does not establish strength across every coding or agent task.
“built for coding and agentic work”
— Meta, in the Muse Spark Open model card, as quoted by ThorstenMeyerAI.com
large memory GPU for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Benchmark Picture
The supplied material does not specify the exact license terms in enough detail to assess all permitted uses. It describes the license as having conditions, so developers should review Meta’s published terms before deploying or redistributing the model. The weights are available, but that alone does not make the model equivalent to a permissively licensed release.
Benchmark scores and provider prices can change as testing methods, model versions, and commercial rates change. The reported cost is for Artificial Analysis’s index task suite, and it may not predict costs for a particular application. The source also does not provide enough detail here to determine how the model performs across all languages, production workloads, or safety evaluations. The 24% hallucination rate is tied to a particular benchmark comparison and should not be read as a general error rate for every user.
As an affiliate, we earn on qualifying purchases.
Check Terms and Workload Costs
Developers considering the model can download the available weights from Hugging Face, review Meta’s model card and license, and test it on representative tasks before committing to deployment. For hosted use, the listed provider rates should be checked directly because prices may differ by provider or change over time.
Further comparisons will depend on additional evaluations of the model across practical workloads, including long-context research, visual tasks, and agentic coding. For now, the available results point to a specific trade-off: strong benchmark performance in selected areas, but a lower overall score than Muse Spark 1.3 and a higher measured task cost than several open-weight alternatives.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
Muse Spark Open is Meta’s 816-billion-parameter mixture-of-experts model, released as downloadable weights on Hugging Face on October 15, 2026. Meta lists text, images, audio or speech, and video as inputs, plus a one-million-token context window.
How did it score against other open-weight models?
Artificial Analysis gave Muse Spark Open 38.9 on its Intelligence Index. The supplied comparison ranks it tenth among open-weight models overall and as the highest-scoring entry from a US lab, while several Chinese-lab models scored higher.
Is Muse Spark Open cheaper than the closed Muse Spark 1.3?
The supplied source reports the same median provider list prices for both models: $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. Downloadable weights do not remove hosted inference charges, and self-hosting requires substantial memory.
What tasks appear to suit the model?
The reported results are strongest on visual reasoning and the AA-Omniscience benchmark, which measures correct answers and penalizes confident errors. Its long context may also suit document or research tasks, though users should test performance on their own material.
Does it perform well on coding agents?
Results are mixed. The source reports 76% on Terminal-Bench 2.1, but 8% on Terminal-Bench 4.0, the harder test used in the Intelligence Index. Those figures do not establish how it will perform in every coding workflow, but they caution against assuming broad agent strength from the model card description alone.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
