Muse Spark Open: Hardware, License Terms and Real Costs for Builders

Muse Spark Open is Meta's 816-billion-parameter mixture-of-experts model. Its weights were released on October 15, 2026, under a custom license that lets most companies use it commercially for free. You need about 700GB of accelerator memory to run it yourself, so most builders will use it through a hosted API at roughly $1.25 per million input tokens and $4.25 per million output tokens. In Meta's own tests it leads on tool use and agent benchmarks but trails the best Chinese open models on pure coding.

This guide covers what the model is, the hardware it needs, what the license really allows, what it costs, and how to decide whether it belongs in your stack.

What Meta released

According to Meta's announcement, Muse Spark Open has 816B total parameters and 42B active per token, a 1M-token context window, and native multimodal input. Meta calls it "the most capable US open-weight model to date." The Hugging Face model card adds the architecture details:

Muse Spark had been a proprietary model since it debuted in April, SiliconANGLE reports. The open version got extra safety post-training so it holds up without Meta's server-side guardrails. Sources disagree on which version it comes from. SiliconANGLE says it was built from Muse Spark 1.2, which shipped in August. The PyTorch blog says it "shares the same base model as Muse Spark 1.3." Hugging Face metadata also lists 822B parameters, while the model's name and card say 816B. None of this affects how you use it, but don't assume it behaves exactly like any hosted Muse Spark version.

Meta's top models stay closed. The newer Muse Wonder models, introduced October 12, are available only through the Meta Model API and Muse Code, according to SiliconANGLE.

Benchmarks: where it leads and where it trails

Every number below comes from Meta's own table on the model card, so none of it has been verified independently yet. VentureBeat notes that it couldn't independently confirm the Artificial Analysis Intelligence Index score. The table compares Muse Spark Open with the three strongest open-weight rivals Meta listed:

Three patterns stand out:

1. Tool use and orchestration are its strengths. MCP-Atlas covers 220 tools across 36 live Model Context Protocol servers, per SiliconANGLE. The model has to pick which tools to call, and in what order, from a plain-language task. If your product is an agent that chains many API calls, this is the benchmark to watch.
2. On raw coding, it is a step behind the Chinese leaders. On Terminal-Bench 2.1 and DeepSWE it scores 6 to 9 points below Kimi K3. Treat the "coding-optimized" label as relative to other US open models, not to the whole field.
3. It is much harder to jailbreak. A 0.5 on StrongREJECT against 60-plus for its rivals is a big gap. The flip side may be more refusals. On the FORTRESS benchmark, its "ORS" score is 17.4, against 2.7 to 12.8 for the rivals listed, and the card doesn't define ORS. If ORS means over-refusal, as the name suggests, test it on your edge-case prompts before you ship.

Among US open models, the gap is wide. Its Intelligence Index score of 39 compares with 26 for Inkling and 23 for Nemotron 3 Ultra.

Hardware: what self-hosting takes

The model card gives the memory floor: at least 700GB of accelerator memory for NVFP4 and 2TB for BF16. In practice, that means a multi-GPU server node for NVFP4 or a multi-node setup for BF16. A single workstation won't do it.

Meta's main demo ran the model on two M5 Ultra Mac Studios linked with Apple's MLX framework, about 1TB of combined unified memory. In the demos, Muse Code fixed bugs in the Mastodon client Ice Cubes and reviewed DaVinci Resolve footage to crop shots for vertical video, according to Meta. Two caveats from VentureBeat matter here:

• The 512GB Mac Studio configuration doesn't ship until late October, and Apple hasn't announced its price. VentureBeat's reference point is four 256GB M5 Ultra units at $10,799 each, about $43,196 before tax.
• Meta hasn't published a tokens-per-second figure for the Mac setup or any other local setup. The demos show that it works, not how fast it runs.

Software support is broad. Meta lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth, optimized for AMD, Apple, Intel and NVIDIA hardware. The PyTorch team shipped Helion kernels that are 1.03 to 3.95 times faster at the kernel level on NVIDIA GB300. End to end, the gain is much smaller: the PyTorch blog reports 3% more throughput on AMD MI350. For fine-tuning, TorchTitan supports multi-turn supervised fine-tuning (SFT), LoRA and reinforcement learning (RL) through TitanRL.

Getting started looks like any other OpenAI-compatible model. The model card shows vllm serve "meta-models/Muse-Spark-Open-816B-A42B" exposing an endpoint on port 8000, plus SGLang and Docker Model Runner equivalents. The recommended sampling settings are temperature 1.0, top_p 1.0 and top_k disabled.

The Muse Model License, in plain terms

The weights are not released under an OSI open-source license. They come under Meta's own Muse Model License. Here is what the model card and press coverage say it means for a typical startup:

A few details matter more than they look:

• The attribution must be prominent. The label goes in your product or its documentation, and VentureBeat reads it as applying to the product interface. White-label resellers and agencies should check whether their clients will accept it.
• The $20M threshold counts affiliates. It applies to aggregate gross revenue, not just revenue from the model. A small subsidiary of a large company may cross it on day one.
• Meta claims no ownership of outputs, and the weights come "as is," with no warranty.
• Downloads are gated. The Hugging Face form asks for your full legal name, date of birth and organization name with corporate identifiers, and you can't edit it after submitting. VentureBeat notes Meta has used similar forms on earlier releases such as Llama 4 Scout.

This is not legal advice. If your business is serving models to others, have counsel read the full license text before you build on it.

Pricing: hosted API versus your own hardware

Meta hasn't published a first-party token price for Muse Spark Open. VentureBeat reports Together AI's rates: $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. At launch, Meta listed Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as hosts.

The context window differs by provider. VentureBeat found Groq listing 131,072 tokens and requiring organization approval, while Fireworks lists the full 1,048,576. If you chose this model for whole-repository context, check your provider's limit first.

Here is a rough break-even example. Suppose a product sends 100M input tokens and 20M output tokens a month with no caching. At Together's rates, that is $125 + $85 = $210 a month. Against VentureBeat's roughly $43,000 Mac reference, the hardware would take more than 200 months to pay for itself, and that's before power, maintenance or any throughput limits. The numbers only favor self-hosting if you have:

• data that can't leave your infrastructure,
• very high, steady token volume on GPUs you already own, or
• a fine-tuned derivative that no host will serve for you.

To try your own figures, plug them into the LLM API Cost Calculator. Use the Token Counter to see how far a large codebase actually gets into a 1M-token window. Muse Spark Open uses its own 202,048-token vocabulary, so treat any o200k count as an estimate.

One cost trap: the default reasoning effort is xhigh. The model card recommends high or xhigh for coding and agentic tasks, and low or medium for faster, shorter replies. If you leave the default on for simple classification or chat, you'll pay for reasoning tokens you don't need.

Who should use it, and who should wait

A good fit:

• Agent and MCP-heavy products. The MCP-Atlas and JobBench leads are its strongest case, and the low jailbreak rate helps when agents handle untrusted input.
• Teams that need a US-origin open model. Some procurement and compliance teams prefer a US model to Chinese ones, and on Meta's numbers this is the strongest US option.
• Multimodal pipelines. It is one of few open models in this class that accepts audio and video natively.

Probably not the first choice:

• Pure coding agents chasing top SWE scores. Kimi K3 and GLM 5.3 score higher on Meta's own table.
• Teams that want hassle-free licensing. If attribution or the hosted-service clause is a problem, Muse Glimmer, Meta's 30B model, ships under Apache 2.0 according to VentureBeat, though it is far less capable.
• Small-scale self-hosting. The 700GB memory floor rules out most single machines.

If you're building coding tools on top of models like this, our AI coding assistants directory shows what others in the space are shipping.

What to do now

1. Test it on your own tasks through a hosted API before buying hardware. Run your hardest agent traces at both high and xhigh reasoning effort, and log refusals.
2. Check your provider's context limit and pricing. Hosts differ on both.
3. Classify your deployment against the license. Is this internal use, a feature in your own product, or a hosted service? Plan for the attribution label now.
4. Watch for independent evaluations. Meta has said a full safety evaluation report is coming. Until third parties such as Artificial Analysis publish their own scores, treat the benchmark table as Meta's claims.
5. Watch Mac Studio pricing when the 512GB configuration ships in late October. Until then, the "runs on two Macs" claim has no price attached.