What it is
Muse Spark Open is an open-weight model from Meta Superintelligence Labs, released on October 15, 2026. It is the open version of Muse Spark and, per Meta, shares its base model with Muse Spark 1.3, with extra safety post-training so it holds up without Meta's hosted filters. It is a sparse mixture-of-experts model with 816 billion parameters, 42 billion active per token, and a context of 1,048,576 tokens. It reads text, images, video and audio and writes text.
Weights come in BF16 (about 1.65TB) and NVFP4 (about 537GB) from a gated Hugging Face repo, and through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Meta does not serve it on its own Model API. Hosted endpoints include Baseten, Fireworks, Together, Nebius and Modal, all listed on OpenRouter at $1.25 input and $4.25 output per million tokens.
It is released under the Muse Model License, not an open-source licence. Use is free, but products must display "Built with Muse Spark Open", and companies selling model access as a service with more than $20 million in annual revenue need a separate agreement with Meta.
Our review
Muse Spark Open is the strongest open-weight model from an American lab on the evidence available, and still not the strongest open model. Artificial Analysis scores it 39 on its Intelligence Index, against 45 for GLM 5.3 and 44 for Kimi K3 on the same scale, as reported on Meta's own model card. Where it leads is agent work: 89.2 on MCP-Atlas against Kimi K3's 82.3 and 61.8 on JobBench against 54.3. On Terminal-Bench 2.1 it trails both, 76.4 against 83.9 and 85.0.
The safety table is the distinctive part. On the adaptive StrongREJECT v2 jailbreak test Meta reports 0.5 against scores of 62.9 to 86.4 for the other open models it tested, and on its cyber jailbreak sets it refused 99.2% and 94.8% where Kimi K3 refused 75.5% and 70.2%. These are Meta's figures, and the card warns they should not be assumed to survive fine-tuning. For a regulated buyer choosing an open model to run internally, they are still the most relevant numbers on the page.
Hosting is where the economics wobble. Every OpenRouter endpoint charges $1.25 and $4.25 per million tokens, exactly the closed Muse Spark 1.3 price, and Artificial Analysis calls it particularly expensive against similar open models, whose class median is $0.43 and $1.59. GLM 5.3 scores higher on that index at $1.40 and $4.40 from Z.ai. The open weights matter for data control, not for price.
Self-hosting is a data-centre job. Meta's smallest official build, NVFP4, is about 537GB of weights and wants a four-GPU Blackwell node, and the model card advises at least 700GB of VRAM once cache memory is counted. The headline desktop option is two 512GB Mac Studios on MLX. The licence is permissive for most users, with no user cap and no claim on outputs, but the gated download asks for a legal name, date of birth and organisation, and every product built on it must carry Meta's attribution line.
Where it earns its place
An infrastructure team at a bank with an on-premises Blackwell cluster
Run an internal coding and agent model where no prompt leaves the building
The NVFP4 build fits a single four-GPU Blackwell node, and every prompt and response stays inside the bank's network. Internal use falls outside the licence's $20M model-as-a-service clause, and Meta's jailbreak results are the strongest of the open models it tested.
A product manager at a mid-sized SaaS company
Ship an agent feature on open weights to avoid lock-in to one API
Start on any OpenRouter endpoint at $1.25/$4.25 per million tokens and move to self-hosting later without changing models. The product must show "Built with Muse Spark Open" somewhere in its interface, docs or settings.
An ML engineer at a vertical AI start-up
Fine-tune a frontier-scale model on proprietary workflows
Meta supports SFT and reinforcement learning through PyTorch and TorchTitan and lets licensees distil and redistribute derivatives. Training at 816B parameters needs a multi-node cluster, and Meta's safety results do not carry over to modified models.
Before you commit
- Setup
- Request access on Hugging Face (legal name, date of birth, organisation), then serve with SGLang or vLLM; or call a hosted endpoint on OpenRouter, Together, Fireworks, Baseten, Nebius or Modal.
- Integrations
- SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth; SageMaker AI, Google Cloud Model Garden and Microsoft Foundry catalogues. Reasoning effort is low, medium, high or xhigh, defaulting to xhigh.
- Data and privacy
- Muse Model License: free use, modification, distillation and redistribution; attribution required in each product; model-as-a-service providers above $20M annual revenue need a separate agreement. Meta claims no ownership of outputs. Self-hosted, nothing leaves your infrastructure.
- Cost at scale
- Self-hosting: NVFP4 (about 537GB) on one four-GPU Blackwell node, BF16 (about 1.65TB) on eight B300s, or two 512GB M5 Ultra Mac Studios. Hosted, a million support replies a month at 3,000 input and 500 output tokens is about $5,875 at OpenRouter rates.
- Limits
- 1,048,576-token context, though one OpenRouter endpoint caps it at 1M and output caps vary by provider; not available through OpenRouter's EU in-region routing. Meta's benchmarks are unverified independently, and the card notes weak handling of very short or low-resolution video.
What's good
- 816B MoE with 42B active and a 1M-token context, open weights
- 89.2 on MCP-Atlas, ahead of Kimi K3, GLM 5.3 and Qwen3.8 on Meta's table
- 0.5 on StrongREJECT v2 against 62.9 to 86.4 for rival open models, per Meta
- Official NVFP4 build; SGLang, vLLM, TensorRT-LLM, llama.cpp and MLX support at launch
What's not
- Artificial Analysis index of 39, behind GLM 5.3 (45) and Kimi K3 (44)
- Hosted price matches closed Muse Spark 1.3, well above open-model medians
- Custom licence: mandatory attribution and a $20M threshold for model-as-a-service
- Gated download asks for legal name, date of birth and organisation
Best for
Organisations that need a capable agent model on their own hardware, with stronger refusal behaviour than other open models, and can live with Meta's attribution clause.
Alternatives
Muse Spark Open: common questions
- Is Muse Spark Open worth it?
- A serious open-weight agent model with unusually strong safety results, held back by hosted prices that match closed Spark and a licence that asks for your birth date. We rate it 4.0 out of 5.
- How much does Muse Spark Open cost?
- Free weights under the Muse Model License (attribution required; $20M revenue threshold for model-as-a-service). Hosted on OpenRouter at $1.25 input, $0.15 cached, $4.25 output per 1M tokens.
- What are the downsides of Muse Spark Open?
- Artificial Analysis index of 39, behind GLM 5.3 (45) and Kimi K3 (44); hosted price matches closed Muse Spark 1.3, well above open-model medians; custom licence: mandatory attribution and a $20M threshold for model-as-a-service; gated download asks for legal name, date of birth and organisation.
- Who is Muse Spark Open best for?
- Organisations that need a capable agent model on their own hardware, with stronger refusal behaviour than other open models, and can live with Meta's attribution clause.
- How much does Muse Spark Open cost for a team or in production?
- Self-hosting: NVFP4 (about 537GB) on one four-GPU Blackwell node, BF16 (about 1.65TB) on eight B300s, or two 512GB M5 Ultra Mac Studios. Hosted, a million support replies a month at 3,000 input and 500 output tokens is about $5,875 at OpenRouter rates.
- Is Muse Spark Open safe for company data?
- Muse Model License: free use, modification, distillation and redistribution; attribution required in each product; model-as-a-service providers above $20M annual revenue need a separate agreement. Meta claims no ownership of outputs. Self-hosted, nothing leaves your infrastructure.
- What are the best Muse Spark Open alternatives?
- The closest alternatives we have reviewed are GLM-5.3, Kimi K3, Muse Spark 1.3 and Muse Glimmer.
Answers checked Oct 23, 2026.
Make Muse Spark Open? Show this rating on your site
