Meta released the weights of Muse Spark Open on 2026-10-15, calling it "the most capable US open-weights model to date."[1][4] This Model Watch reads Meta's announcement, the Hugging Face model card, Meta's developer blog and the PyTorch post to answer three practical questions: what the licence lets you do, what it takes to run an 816B-parameter model, and how much the benchmark table can be trusted.
Everything below comes from those four pages, read on 2026-10-18. We did not run the model, so nothing here is our own testing. We also could not open the full licence file (its repository path returned a 404), so licence terms are those quoted on the model card; read the licence itself before you build on it. Meta's evaluation methodology report is linked from its announcement, but we did not open it.
Quick answer
- What it is. A mixture-of-experts transformer with 816B total parameters and 42B active per token, a 1M-token context window, text, image, video and audio input, and text-only output.[2]
- Licence. The Muse Model License: attribution notice, a separate agreement with Meta above $20 million gross revenue for model-as-a-service offerings, and an acceptable use policy. Fine-tuning, distillation and merging are permitted.[2]
- Hardware. Not a single-GPU model. The smallest official recipe is a Blackwell Ultra or H200 cluster, or two to four high-memory Mac Studios.[3]
- Benchmarks. Self-reported by Meta at the highest reasoning effort. The card lists open-weight rivals that score higher on some rows.[2]
- Pricing. Meta's developer page lists no prices; hosts such as Fireworks AI and Modal sell per-token access, so check each provider.[3]
Specifications
| Spec | Muse Spark Open (model card) |
|---|---|
| Parameters | 816B total, 42B active per token (the Hugging Face metadata field lists 822B) |
| Architecture | Mixture-of-experts transformer with a 1.8B-parameter Perception Encoder |
| Layers and experts | 62 layers, 256 experts, 8 active per token |
| Attention | 128 query heads, 16 key-value heads (grouped-query attention), gated attention with QK-norm |
| Context | 1M tokens at inference |
| Input and output | Text, image, video, audio in; text out |
| Languages | Trained on data from more than 100 languages |
| Reasoning effort | low, medium, high, xhigh (default xhigh) |
| Released weights | BF16 and NVFP4, Safetensors format |
The 16 key-value heads against 128 query heads matter for serving: fewer key-value heads shrink the KV cache that grows with every token of that 1M-token window. We explain the mechanism in Inference Serving Explained. The 42B active figure is why cost per token resembles a mid-sized model, while the 816B total is why memory does not.[2]
Meta says the model shares a base model with Muse Spark 1.3, has additional safety post-training for use "as a standalone model without system guardrails," and is trained for coding and agentic work including subagent coordination and computer use.[4] Those are capability claims from the vendor.
The licence, term by term
The model card quotes these terms of the Muse Model License ("Muse Spark Open" version dated 2026-10-15):[2]
| Term | What the card quotes |
|---|---|
| Attribution | The licensee "shall prominently display 'Built with Muse Spark Open'" in the product interface, documentation, website or settings menu |
| Revenue threshold | A licensee and its affiliates that offer model access as a service and exceed $20 million aggregate gross revenue in any consecutive 12 months "must notify Meta and enter into a separate written agreement with Meta before offering the Software"; the card says internal use, research and embedding the model in your own product are excluded |
| Outputs | "Meta claims no ownership of outputs generated by the Software" |
| Derivatives | Fine-tuning, distillation and merging are permitted |
| Acceptable use | Exhibit A prohibits illegal activity, harassment, unlicensed professional practice, malware, weapons and self-harm content, deception and disinformation, and circumventing safety measures; not intended for users under 18 |
| Warranty | Provided "as is" |
Meta's announcement describes these as "limited commercial terms, which allow many developers to use the model for free."[1]
Our interpretation
Based on the quoted terms, the practical split is this:
- Most teams are unaffected by the revenue clause. It targets companies that resell model access as a service at scale. Using the model inside your own product is described as excluded.
- Attribution is a product requirement, not a formality. If you ship it in a customer-facing app, someone must add the notice to the UI, docs or settings.
- The acceptable use policy travels with the weights. That is a usage restriction, and it is the reason the licence differs from permissive licences such as Apache 2.0.
Meta itself uses the term "open-weights" rather than open source in its announcement, while the PyTorch post also uses the phrase "open source" loosely in a safety claim.[1][4] Whether the licence satisfies the Open Source Initiative definition is a separate question. Our guide to open weights versus open source sets out the test, and a licence with field-of-use restrictions and a commercial agreement clause is the kind of case it addresses. We have not run that test on this licence because we have not read its full text.
What it takes to run
| Format | Figure | Source |
|---|---|---|
| NVFP4 | "at least 700GB of VRAM" | Model card[2] |
| NVFP4 checkpoint | about 530 GB, only experts quantised; four Blackwell Ultra GPUs or eight H200s | Developer blog[3] |
| BF16 | about 2 TB | Model card[2] |
| BF16 reference weights | about 1.6 TB | Developer blog[3] |
| Apple | Two M5 Ultra Mac Studios with 512 GB each, or four with 256 GB each (1 TB total), using MLX | Developer blog[3] |
| Community GGUF | Quants down to about 2-bit (Unsloth); 20 community GGUF models listed | Developer blog, model card[2][3] |
The numbers do not all agree, and the reasons are partly arithmetic. 816B parameters at two bytes each is about 1.6 TB, which matches the developer blog's BF16 figure; the model card's roughly 2 TB likely includes more than weights, but the card does not say. Likewise a 530 GB checkpoint and a 700 GB VRAM requirement are different quantities: the second plausibly adds working memory such as the KV cache. These are our readings, not statements from Meta. The developer blog also gives two different smallest local footprints for the lowest-bit quants (about 300 GB and a 256 GB setup), so plan from a measured run, not from either figure.[3]
Serving engines named on the pages are SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, MLX and Unsloth. For fine-tuning, Meta points to TorchTitan (supervised fine-tuning, LoRA and reinforcement learning), among other tools.[2][3][4] Speculative decoding drafters are provided; Meta reports 3.7 times single-request throughput on AMD MI355X with accuracy unchanged, a vendor figure for one hardware setup.[3]
If you only need the output, seven hosted providers are named: Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter and Together AI.[3] That route avoids the memory bill but replaces the licence question with each host's terms.
The benchmark table
The model card reports results at xhigh reasoning effort, all produced by Meta:[2]
| Benchmark | Score |
|---|---|
| GPQA Diamond | 89.0 |
| OSWorld Verified | 80.4 |
| MCP-Atlas | 89.2 |
| Terminal-Bench 2.1 | 76.4 |
| Toolathlon Verified | 72.8 |
| DeepSWE 1.1 | 60 |
| Humanity's Last Exam, no tools | 26.1 |
| Humanity's Last Exam, with tools | 61.0 |
| Artificial Analysis Intelligence Index | 39 |
How to read it:
- These are vendor numbers. The card does not identify an independent reproduction, and the announcement's own methodology report was not available to us.
- The card is candid about rivals. It compares against other open-weight models including Kimi K3, GLM 5.3 and Qwen3.8, and on several agentic and multimodal benchmarks some of those score higher.[2] "Most capable" is therefore a claim about the overall picture, not every row.
- Effort setting changes the cost. The scores are at xhigh, the default. Higher reasoning effort uses more thinking tokens; see our guide to reasoning models, thinking tokens and cost. Expect different latency and cost at lower settings.
- Safety evidence is promised, not shown. Meta says the model met its release thresholds under its scaling framework and that a full evaluation report is "available soon."[1] The card lists one safety figure (StrongREJECT v2, adaptive, lower is better: 0.5).[2] Until the report is published, treat the safety claim as unverified.
For how to weigh a single vendor table, see How to Read an AI Benchmark Claim. For Meta's separate hosted model and its benchmark rules, see our Muse Wonder Model Watch.
Who should consider it
- Teams with data-residency or customisation needs and a GPU budget. The permitted fine-tuning and distillation, plus TorchTitan support, are the real draw. The hardware floor is high.
- Teams that want open weights without hosting. Use a named provider and compare price against closed models on your own evals.
- Hobbyists. The community quantisations exist, but this is a 42B-active, hundreds-of-gigabytes model. Smaller open models are a more realistic local choice.
- Anyone who needs a permissive licence. The attribution notice and acceptable use policy rule that out.
What is still uncertain
- The full licence text and any updates after 2026-10-15.
- Independent benchmark reproductions, and Meta's full safety evaluation report.
- Real-world memory and throughput on your hardware.
- Prices, which Meta does not publish and which vary by host.
