🔍 Read the full analysis: Muse Spark Open Brings America’s Strongest Open-Weight Model To Market 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 its weights available on Hugging Face. Artificial Analysis scored it 38.9, the highest result for a US lab’s open-weight model, while finding strengths in hallucination resistance and visual reasoning. The model’s licensed weights are free, but provider prices match Meta’s closed Muse Spark 1.3, and several open rivals score higher for less.
Meta released Muse Spark Open, an 816-billion-parameter model, on October 15, publishing its weights on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index v4.3.2, the highest score among open-weight models from US labs in that index; the result places it tenth overall, behind nine models from Chinese labs.
The model uses a mixture-of-experts design with 256 experts and 42 billion active parameters per token. It accepts text, images, speech and video as input, produces text, and supports a one-million-token context window. Meta provides BF16 and NVFP4 weights. Its model card estimates self-hosting requires about 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 Muse Spark Open at its default xhigh reasoning effort. Its 38.9 score is rounded to 39 in the index. The result is well above Meta’s earlier open releases cited in the source: Muse Glimmer scored 17 and Llama 4 Maverick scored 9. NVIDIA’s Nemotron 3.5 Super VL scored 26.5 in a same-day release, according to the source material.
The index also records uneven results across tasks. Muse Spark Open scored 24.7 on AA-Omniscience, highest among 167 open-weight entries, with a 24% hallucination rate among the 15 highest-scoring open models. It scored 81% on MMMU-Pro visual reasoning. On Terminal-Bench 4.0, however, Artificial Analysis measured 8%; its score on Terminal-Bench 2.1 was 76%. These are benchmark results, not a guarantee of performance on a particular workload.
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
A US Model With a Cost Trade-Off
Muse Spark Open gives US developers access to a substantially larger open-weight model from Meta’s flagship Muse Spark line. Its visual reasoning, long context and lower measured hallucination rate may suit document-heavy research tasks. The model’s tenth-place position overall also shows that a high US ranking does not make it the leader across the full open-weight field.
Access and operating cost are separate questions. The weights are available without a charge, but the source reports that four third-party providers list a median price of $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million. Those rates match Meta’s closed Muse Spark 1.3 pricing. Artificial Analysis estimated $0.84 for a weighted index task, compared with $0.13 for MiMo-V2.6-Pro and $0.27 for DeepSeek V4.1 Flash.
That makes workload choice important. A team may value the model’s visual and long-context results, but should weigh them against hosted inference costs and task-specific performance. In the source’s index comparison, Muse Spark 1.3 scored 45.1 at the same reasoning effort and list price, while several open models scored more highly at lower estimated task costs.
high-performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Lineup
Muse Spark Open is the first open-weight release in Meta’s Muse Spark family, according to the source. It follows smaller open models, including Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0. The new model changes the scale of Meta’s open offering, though its reported index score remains below the closed Muse Spark 1.3 score of 45.1 and Muse Wonder score of 52.4.
The wider comparison is led by models from Chinese labs. Artificial Analysis scored Xiaomi’s MIT-licensed MiMo-V2.6-Pro at 46.3. The source lists eight other models ahead of Muse Spark Open, including GLM-5.3, Step 5 Preview, Kimi K3, two Qwen3.8 models and DeepSeek V4.1 Flash. Licensing terms vary, so score and price alone do not describe the full differences in how developers can use each model.
“built for coding and agentic work”
— Meta, in the 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.
Workload Fit and Licensing Terms
The benchmark results do not establish how Muse Spark Open will perform across real production workloads. The source gives contrasting agent scores: 76% on Terminal-Bench 2.1 and 8% on the harder Terminal-Bench 4.0. It is not clear from the supplied material how those differences will translate to specific coding or agent deployments.
The source also does not provide the full text of the model’s license or explain all usage conditions. Developers should review Meta’s published terms before building on the weights. Provider prices are listed rates and may vary by provider or over time; the supplied material does not say how broadly those rates or benchmark costs will hold in practice.
As an affiliate, we earn on qualifying purchases.
Provider Access and Developer Testing
The weights are already listed on Hugging Face, and the source reports that four third-party providers offer API access. The next practical steps for developers are to review Meta’s license, compare provider pricing for their expected input and output mix, and test the model on representative tasks. Further provider availability, real-world operating costs and performance reports may clarify where the model is most useful.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released the model and published its weights on Hugging Face on October 15, 2026, according to the source material.
How did Muse Spark Open score?
Artificial Analysis scored it 38.9 on Intelligence Index v4.3.2 at its default xhigh reasoning effort. The index rounds that to 39, placing it tenth among open-weight models overall and first among those from US labs.
Are the weights and hosted use both free?
The source says the weights are free to access. Hosted API use is priced by providers; their reported median list price matches Meta’s closed Muse Spark 1.3 rates: $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million.
What tasks appear to be strengths?
Artificial Analysis reported a 24.7 AA-Omniscience score, a 24% hallucination rate among the top 15 open-weight models, and 81% on MMMU-Pro visual reasoning. These are benchmark findings and may not predict results for every use case.
Is it a strong model for terminal agents?
The supplied results are mixed. Artificial Analysis measured 8% on Terminal-Bench 4.0, compared with 76% on Terminal-Bench 2.1. Developers considering terminal agents should evaluate the model on their own tasks and account for hosted inference costs.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
