Meta releases Muse Spark Open: 816B open weights, a 1M-token window, and a license with a revenue clause

On October 15 Meta released Muse Spark Open, the open-weight version of its Muse Spark model. The weights are on Hugging Face under a custom Muse Model License dated the same day, and the model is already deployable on Amazon SageMaker JumpStart. It arrives three days after Meta's closed frontier model, Muse Wonder, reached the Meta Model API, and it fills a different slot: a very large model you can run on your own infrastructure.

Meta calls it the most capable open-weight model from a U.S. company. That ranking rests on evaluation results Meta published itself, as TechCrunch notes; no independent evaluation was available at the time of writing.

What was released

According to the Hugging Face launch post, written with the integration teams that support the model, Muse Spark Open is a mixture-of-experts model with about 816 billion total parameters and about 42 billion active per token. Each MoE layer has 256 experts and routes every token to 8 of them. It accepts text, images, audio and video, and produces text. Meta trained it mainly for agentic coding: reading a codebase, multi-file refactors, long debugging sessions and tool calls.

• Context window: up to 1M tokens (1,048,576 positions). Image, video and audio tokens count against the same window as text.
• Checkpoints: BF16 (about 1.5 TB in safetensors) and NVFP4, a 4-bit format aimed at NVIDIA Blackwell GPUs.
• Reasoning: always on. The chat template exposes low, medium, high and xhigh effort levels, defaulting to xhigh; there is no way to switch reasoning off.
• Knowledge cutoff: the default chat template declares January 4, 2026.
• Family: it shares a tokenizer with Muse Glimmer, the small dense model Meta released in August.

What it takes to run

Only 42B parameters compute each token, but all 816B have to sit in memory, so hardware is sized by memory, not by compute. The Hugging Face post gives concrete reference points:

• In Transformers, the BF16 checkpoint runs on four nodes of eight H200 GPUs; the NVFP4 checkpoint fits a single B200 node.
• On Hugging Face Inference Endpoints, the NVFP4 recipe runs on 8× RTX Pro 6000 with a 131,072-token context, or on 8× H200 for the full 1M tokens.
• On a Mac Studio with M3 Ultra and 512 GB of memory, a 2-bit llama.cpp quant used 304 GB and produced about 18 tokens per second; a 3-bit MLX quant used about 344 GB and generated about 21 tokens per second in a short test. Those are the Hugging Face team's own measurements.
• On SageMaker JumpStart, AWS lists the NVFP4 variant on ml.p5e.48xlarge and BF16 on ml.p6-b300.48xlarge.

The long context has a real memory cost of its own. Using the model's published attention configuration, the Hugging Face post estimates about 250 KB of KV cache per token in BF16, or roughly 266 GB for a full 1M-token context, on top of the weights. The headline "runs on a Mac" therefore means a top-end Mac with an aggressive quantization and a much shorter working context, not a laptop.

Beyond self-hosting, TechCrunch reports the model is available now through a number of API platforms. Pricing for hosted inference is set by each provider and was not part of Meta's materials that we could open.

The license, read closely

The Muse Model License is permissive in most respects: use, modification, fine-tuning, distillation and redistribution are allowed at no cost. Three conditions matter for developers:

• Products built on the model or its derivatives must display "Built with Muse Spark Open" somewhere in the UI, documentation or settings.
• Companies selling access to the model as a service (via API or similar) with more than $20 million in gross revenue over 12 months must sign a separate agreement with Meta first. Internal use, research and using the model as a component of your own product are exempt.
• Use is bound by an attached usage policy, which also states the model is not intended for people under 18. Download on Hugging Face requires a manually approved access request.

That is open weights, not open source in the OSI sense. For most teams the revenue clause will not bite; for inference providers and resellers, it will.

What changes for developers

For anyone choosing a model for coding agents, Muse Spark Open adds a large, agent-oriented option that can run inside your own network. AWS's pitch is exactly that: the endpoint lives in your account, so source code and prompts stay on infrastructure you control, and the model version only changes when you redeploy. That matters for regulated codebases and for teams that want reproducible agent behavior.

Three practical notes. First, because reasoning cannot be disabled, simple tasks still pay for thinking tokens; the Hugging Face post recommends the low setting for easy work. Second, the 1M window is a ceiling, not a default: most practical deployments will run far shorter contexts unless they budget hundreds of gigabytes for KV cache, so retrieval and context trimming still matter. Third, the chat template follows the channel-based Harmony format first used by gpt-oss, which separates reasoning, commentary and tool calls; harnesses that already handle that format should need fewer changes.

What is still missing is outside evidence. Meta's comparisons with other open-weight models are its own, and the quantized speeds above come from short tests on single machines. Until independent benchmarks and longer agentic evaluations appear, treat the capability claims as the vendor's.