Muse Spark Open

What Meta released

Muse Spark Open is the open-weight member of Meta’s Muse Spark model family, released on October 15, 2026 by Meta Superintelligence Labs. Meta describes it as the most capable US open-weight model to date and says it was built for coding, agentic work, long-horizon tasks, and multi-step reasoning. That “most capable” line is Meta’s own framing; what you can verify is the model card, the license, and how it performs on your work.

It is a separate thing from the other Muse products. Meta Muse is the consumer agent, Muse Code is a paid coding agent, and Muse Glimmer is a much smaller open model meant for a single GPU. Muse Spark Open is a large set of weights you can download and run, or call through a hosting provider.

Total parameters versus active parameters

The headline numbers are 816 billion total parameters and 42 billion active. Those two figures measure different things, and mixing them up leads to wrong conclusions.

Muse Spark Open is a Mixture-of-Experts model. Each MoE layer contains 256 smaller “expert” networks, and a router picks 8 of them for each token. So while the model stores 816B parameters of knowledge, only about 42B do work on any given token. That keeps the compute per token closer to a mid-sized dense model.

The catch is memory. Even though only 42B parameters are active at a time, the router can pick any expert, so the whole 816B checkpoint has to be loaded. Active parameters tell you roughly how fast and how expensive each token is to compute. Total parameters tell you how much hardware you need to hold the model. Neither number, on its own, tells you how good the model is.

Specs

What the license allows

The weights come under the Muse Model License, not Apache 2.0 or MIT. In practice:

• You must prominently display “Built with Muse Spark Open” in your product, docs, site, or settings, including for fine-tuned derivatives.
• If you offer the model to third parties as an API or similar service and your aggregate revenue tops $20 million in any 12-month period, you must notify Meta and sign a separate agreement first. Internal use, research, and using it inside your own product are excluded from that clause.
• Meta claims no ownership of outputs.
• An acceptable use policy prohibits a long list of uses, and the model is not intended for users under 18.

Downloading requires a form with your full legal name, date of birth, and organization, which cannot be edited after submission. That is a meaningful difference from models you can pull anonymously. Our open weights entry explains why “open-weight” and “open source” are not the same thing.

Running it

Self-hosting is serious infrastructure. Meta’s card lists at least 700GB of GPU memory for NVFP4 and about 2TB for BF16. Meta also says the model is optimized for two Mac Studios with M5 Ultra using MLX, which is an unusual and comparatively accessible option for a model this size. Supported frameworks listed on the card include SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth, and llama.cpp, and community quantizations exist for Ollama and LM Studio. Heavier quantization lowers the hardware bar at some cost in quality.

Hosted access is easier. Meta’s developer docs list Baseten, Fireworks AI, Groq, Modal, Nebius, and Together AI as launch inference platforms, OpenRouter as a gateway, and Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry as catalogs. Each provider sets its own price and data terms. Meta does not offer this model through its own API with a published price.

Benchmarks, read carefully

Meta’s model card reports, among others, 76.4 on Terminal-Bench 2.1, 80.4 on OSWorld Verified, 89.2 on MCP-Atlas, 60 on DeepSWE 1.1, and 89.0 on GPQA Diamond, all at the xhigh reasoning setting. The card also shows an Artificial Analysis Intelligence Index of 39.

These come from Meta’s own model card, with Meta choosing the benchmarks, settings, and comparison models. Some are well-known public tests and some are newer agent benchmarks. Independent evaluations were still thin in the first week. Use the numbers to decide whether the model is worth testing, not as proof it beats something else on your workload. High reasoning effort also means more output tokens per task, which affects your real cost; see cost per completed task.

Limitations

Meta’s card notes the model can still err in multi-step reasoning, especially in novel situations, and has not been evaluated in all of its pre-training languages. Quantized versions may differ slightly at the edges. Very short or low-resolution videos may be refused; Meta recommends clips at least 448 pixels on the short side. Meta also recommends human confirmation for irreversible agent actions and not exposing raw reasoning traces to users. The hardware requirement is the biggest practical limit for anyone hoping to run it locally.

How it compares conceptually

Muse Spark Open lands in the same tier as other large open-weight MoE models aimed at agents: Mistral Large 4 (around 1T total, 52B active), DeepSeek V4-Pro (1.6T total, about 49B active), GLM 5.2, Qwen3.8-Max, and Kimi K3. Its distinguishing points are native audio and video input, a US origin that matters to some procurement teams, and a license with a $20M service threshold rather than a fully permissive one. For smaller hardware, MiniMax M3.1 Flash and DeepSeek V4.1 Flash activate far fewer parameters.

For the bigger picture, see Frontier Open Models Explained and Open Models vs Closed Models.

Compare: Muse Spark Open vs Mistral Large 4 and Muse Spark Open vs DeepSeek V4.1 Flash.