🔍 Read the full analysis: Muse Spark Open: A Strong Open-Weight Model With A Closed-Model Cost on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Spark Open on October 15, 2026, placing its weights on Hugging Face. Artificial Analysis scored it 38.9, but its measured cost per Intelligence Index task was $0.84, the same per-token list price as closed Muse Spark 1.3, which scored higher. The model stands out for hallucination and visual-reasoning results; its Terminal-Bench 4.0 score was weak.
Meta released Muse Spark Open on October 15, making the weights of its 816-billion-parameter model available on Hugging Face. Artificial Analysis scored it 38.9 on its Intelligence Index, but measured a cost of $0.84 per benchmark task; hosted access costs the same per token as Meta’s closed Muse Spark 1.3, which scored higher.
Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token, a 1-million-token context window, and support for text, images, speech and video as inputs. Its outputs are text. The weights are offered in BF16 and NVFP4 formats. Meta’s 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 the model at its default xhigh reasoning effort. Its 38.9 Intelligence Index v4.3.2 result is rounded to 39 in the published summary. The source report ranks it tenth among open-weight models overall and says the nine models ahead of it are from Chinese labs. It describes Muse Spark Open as the highest-scoring open-weight model from a US lab, while noting that NVIDIA’s same-day Nemotron 3.5 Super VL scored 26.5.
The model’s list prices from four third-party providers had a reported median of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Those are the stated list prices for closed Muse Spark 1.3 as well. Weights are available without charge, but using a provider still incurs token fees. On Artificial Analysis’ weighted index run, the model cost $0.84 per task; the report attributes $0.67 of that to input, including cache reads and writes.
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 Accuracy, Uneven Practical Value
The release gives developers access to a large model with notably strong results on some measured tasks. Artificial Analysis recorded a 24% hallucination rate, the lowest among the 15 highest-scoring open-weight models in its comparison. It also scored 24.7 on AA-Omniscience, which the source says was first among 167 open-weight entries, and 81% on MMMU-Pro visual reasoning, first among 66 tested open-weight models. Those results may make it worth evaluating for research and document tasks, particularly where visual input and long context matter.
But the headline index score does not make the model a general-purpose bargain. Closed Muse Spark 1.3 scored 45.1 at the same stated list price, while Xiaomi’s MIT-licensed MiMo-V2.6-Pro scored 46.3 and had a reported benchmark task cost of $0.13. DeepSeek V4.1 Flash scored 39.5 for $0.27 per task. These are results from the cited evaluation and pricing snapshot, not a guarantee of costs or performance in a particular deployment.
Agentic use is a particular caution. Artificial Analysis measured 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. The source report also notes a 50% AutomationBench-AA result and lower GDPval-AA and AA-Briefcase scores than selected competitors. For teams, that means the decision is not simply whether the model is open: task reliability, operating cost, hardware and license terms all need checking.
As an affiliate, we earn on qualifying purchases.
Meta’s Open Model Moves Upmarket
Muse Spark Open is the first open-weight release in Meta’s flagship Muse Spark line, according to the source report. It is a marked change in scale from the company’s recent open releases. Muse Glimmer, a 30-billion-parameter model released in August 2026 under Apache 2.0, scored 17 on the current Artificial Analysis index cited by the report. Llama 4 Maverick, released in April 2025, scored 9.
The comparison also shows why the release is strategically notable without making it the overall open-model leader. The source places several Chinese models above it on the index, including MiMo-V2.6-Pro at 46.3. It describes Muse Spark Open as a substantial improvement over other US open-weight offerings but says it does not change the top of the overall ranking. NVIDIA’s Nemotron 3.5 Super VL, released the same day, scored 26.5 in the same source comparison.
Access to weights and access to affordable inference are separate questions. Meta’s model card gives substantial memory estimates for self-hosting, while third-party hosting charges the listed per-token rates. The model’s openness can allow organizations to inspect and run the weights under the applicable terms, but it does not make the hardware or hosted use cost-free.
As an affiliate, we earn on qualifying purchases.
Deployment Costs and License Terms
The supplied source does not give the full text of Muse Spark Open’s license or explain all conditions on commercial use, redistribution, fine-tuning or deployment. Those details matter to organizations deciding whether the weights meet their needs; users should review Meta’s license and model card directly rather than infer permissive terms from the word “open.”
Pricing figures are a reported snapshot of median list prices from four third-party providers. Actual bills can vary with provider, usage pattern, discounts, caching and workload. The source also does not establish that benchmark costs will predict the cost of a production task. Nor does it provide independent evidence that the reported self-hosting configuration will be economical for a given team.
Benchmark rankings reflect particular tests and settings, including the stated xhigh reasoning effort. Results do not settle how the model performs across all coding, research or agent workflows. The source gives some strong and weak benchmark results, but the reasons for the gap between Terminal-Bench versions and performance in real deployments remain unclear.
As an affiliate, we earn on qualifying purchases.
Check Terms Before Deployment
Developers can download the BF16 or NVFP4 weights from Hugging Face and compare the model on their own workloads. Before deployment, organizations will need to confirm the license conditions, hardware requirements, provider rates and whether caching changes their expected input costs. For shell-based or unattended agents, the reported Terminal-Bench 4.0 result is a reason to test carefully rather than assume the model card’s broad agentic positioning applies to a specific workflow.
Further provider pricing, independent evaluations and practical deployment reports could clarify whether the model’s accuracy and reduced hallucination rate justify its serving costs. Until then, the available evidence supports a measured conclusion: Muse Spark Open is a significant high-capability US open-weight release, but its value depends on the task, deployment method and price comparison—not its benchmark score alone.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is Meta’s 816-billion-parameter mixture-of-experts model, with 42 billion active parameters per token. Meta released its weights on Hugging Face on October 15, 2026.
How did it score?
Artificial Analysis recorded 38.9 on its Intelligence Index v4.3.2 at the default xhigh reasoning effort. The source report says that puts it tenth among open-weight models overall and first among US open-weight models in that comparison.
Is it free to use?
The weights are available without charge, but hosting them requires substantial hardware, and third-party inference is billed by tokens. The reported median list prices were $1.25 per million input tokens and $4.25 per million output tokens, with cached input listed at $0.15.
Is Muse Spark Open better than Muse Spark 1.3?
Not on the cited Intelligence Index comparison: closed Muse Spark 1.3 scored 45.1, against 38.9 for Muse Spark Open. The source says both have the same listed per-token prices, though results and costs can differ by provider and workload.
What remains uncertain?
The supplied source does not spell out the full license conditions or establish real-world costs and performance across deployments. Teams should check the license, hardware needs and provider pricing, then test the model against their own tasks.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
