Muse Spark Open: Meta's 816B Open Model, Explained

Meta has spent 2026 selling its AI. Muse, the personal agent it launched in September, runs on closed models, and Muse Wonder, announced on 12 October, is a paid line. Then on 15 October it gave one away. The new release puts the weights of a large, recent Meta model on Hugging Face, and Meta calls it "the most capable US open-weights model to date." Here is what it is, what it needs, and where it falls short.

What Muse Spark Open is

It is a mixture-of-experts model with 816 billion parameters, of which 42 billion work on each token: 256 experts, 8 active at a time, according to the model card. It reads up to 1 million tokens of context, accepts text, images, video and audio, and writes text. Meta trained it on data in more than 100 languages, though the card notes it has not been evaluated in all of them.

Meta's PyTorch post says it shares the base model of Muse Spark 1.3, the model behind the Muse agent, with extra safety training so it can be used without Meta's own system guardrails. The card lists its intended jobs as multi-step reasoning, end-to-end agent tasks, coding agents, tool use and multimodal reasoning.

Muse Spark Open benchmarks against other open models

Meta's model card compares it with the leading open-weight models. The picture is mixed, and Meta's own numbers show it:

Benchmark	Meta's open model	Qwen3.8	GLM 5.3	Kimi K3
MCP-Atlas, tool use	89.2	84.5	84.2	82.3
JobBench, agent tasks	61.8	52.7	61.4	54.3
HLE-Diamond, no tools	26.1	17.3	16.4	22.2
AA Intelligence Index	39	40	45	44
DeepSWE 1.1, coding	60	57	67	68
Terminal-Bench 2.1	76.4	82.0	83.9	85.0
GPQA Diamond	89.0	93.5	91.7	93.5

It leads on tool use and agentic work, and on the hardest reasoning test without tools. It trails all three rivals on Terminal-Bench and on Artificial Analysis's overall index, and trails GLM and Kimi on DeepSWE. The Decoder adds that it falls well short of Meta's own paid Muse Wonder Full.

Safety is where it is furthest ahead. On StrongREJECT v2's adaptive jailbreak test, where lower is better, the card reports 0.5 against scores above 60 for Qwen, GLM and Kimi, and a 99.2% refusal rate on Meta's internal cyber-jailbreak benchmark.

What hardware it needs

This is not a laptop model. The card asks for at least 700GB of GPU memory with the NVFP4 weights and 2TB for full BF16. The headline Apple angle is real but narrow: Meta says it runs on two Mac Studios with M5 Ultra chips using Apple's MLX framework, which per The Decoder means two machines with 512GB of unified memory each. Supported servers include SGLang, vLLM, TensorRT-LLM, Unsloth and llama.cpp.

For most people the practical route is an API. Meta's AI account lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as hosts at launch, and the weights are also on Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. Download from Hugging Face is gated: Meta's form asks for your full legal name, date of birth and full organisation name.

The licence: free, with two strings

The Muse Model License lets you use, modify and ship the model, with two conditions that matter for most builders:

• Attribution. Any product or derivative must prominently display "Built with Muse Spark Open", in the interface, documentation, website or settings.
• The $20 million line. If you sell access to the model itself as a service and your group's revenue passes $20 million in any 12 months, you must tell Meta and sign a separate agreement. Internal use, research, and features inside your own product are exempt.

Meta claims no ownership of outputs, the model is not intended for users under 18, and an acceptable-use policy bans uses such as malware, fraud and unlicensed legal, medical or financial advice.

Who should care

• Indian startups that need data to stay in-house. An open model with a 1 million token context and a permissive licence can be self-hosted in an Indian data centre, if you can afford the GPUs.
• Agent builders. The tool-use and JobBench numbers are the best among open models in Meta's table, and the safety scores make it easier to deploy without your own filter stack.
• Coding teams. Look at Kimi K3 or GLM 5.3 first, or stay with a closed model such as Claude; our guide to Gemini 4 Argon compares the closed frontier models on price.

Meta's consumer agent, Muse, is covered in our roundup of OpenAI Dots competitors, and the open-weight DeepSeek models are in our report on DeepSeek's outage.

FAQ

Is Muse Spark Open free?

The weights are free to download and use under the Muse Model License. You must display "Built with Muse Spark Open", and companies selling the model as a service with over $20 million in revenue need a separate agreement with Meta.

Can I run Meta's open model on a Mac?

Only on serious hardware. Meta says it runs on two Mac Studios with M5 Ultra chips using MLX. On GPUs it needs at least 700GB of memory with the NVFP4 weights.

Is it better than Qwen, GLM or Kimi?

At agent and tool-use tasks, by Meta's numbers, yes. At coding and on Artificial Analysis's overall index it trails GLM 5.3 and Kimi K3.

What is the Muse Spark Open context window?

1 million tokens at inference, per Meta's model card.

Where can I use it without downloading the weights?

Through API hosts including OpenRouter, Together AI, Groq, Fireworks AI, Baseten, Modal and Nebius, and in Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry.