Meta opens Muse Spark Open weights: an 816B multimodal MoE with 1M context, licensed with a separate deal for big model-service providers
On October 15, 2026 Meta released Muse Spark Open on Hugging Face, an open-weight model with about 816 billion total and about 42 billion active parameters per token. It takes text, image, audio and video input, outputs text, supports up to 1 million tokens of context and ships in BF16 and NVFP4. It uses a custom Muse Model License that allows commercial use, fine-tuning and distillation but requires “Built with Muse Spark Open” attribution and a separate agreement for Model-as-a-Service providers with more than $20 million in revenue. Meta’s own benchmarks put it roughly level with Qwen, GLM and Kimi open models, ahead on some tests and behind on others.
Meta’s model card dates Muse Spark Open to October 15, 2026 and describes it as the full-size model in the Muse series; Muse Glimmer, a small dense model released in August, shares its tokenizer. A same-day Hugging Face post says the mixture-of-experts model has about 816 billion total parameters with about 42 billion active per token, using 256 experts per MoE layer and routing 8 per token. It accepts text, images, audio and video, outputs text, and supports up to 1 million tokens of context at inference. Meta says it was trained for agentic coding: reading codebases, writing and debugging code and calling tools. The Hugging Face post calls it the most capable US open-weights model to date.
Weights come in BF16 and NVFP4 repositories, and the BF16 checkpoint is about 1.5 TB. According to the model card, running it locally takes at least about 700 GB of GPU memory for NVFP4 and about 2 TB for BF16. Supported frameworks are SGLang, vLLM, Hugging Face transformers, TensorRT-LLM, Unsloth and llama.cpp, it is optimized to run on two Mac Studios with M5 Ultra using MLX, and it is available via API through third-party inference providers. Hugging Face says the BF16 checkpoint runs in transformers on 4 nodes with 32 H200 GPUs in total and NVFP4 on a single B200 node, and that a 2-bit quant ran at about 18 tokens per second on an M3 Ultra Mac with 512 GB of memory. Reasoning cannot be turned off; effort can be set to low, medium, high or xhigh, with xhigh as the default.
The model uses Meta’s custom Muse Model License. It permits free use, modification, fine-tuning, distillation, merging, distribution and sale, but the license and Usage Policy must be included, and products or services using the model or its derivatives must prominently display “Built with Muse Spark Open” in the interface, documentation, website or settings menu. If a licensee and its affiliates offer models to third parties as a service via API or similar and their combined gross revenue exceeds $20 million in any 12-month period, they must notify Meta and sign a separate written agreement first; internal use, research, commercial activity that is not Model-as-a-Service, and using the model to power features of one’s own product are exempt. Meta claims no ownership of outputs. The Usage Policy bans uses such as generating malicious code, the model is not intended for people under 18, and downloading it from Hugging Face requires submitting a full legal name, date of birth and full organization name.
In Meta’s own comparison table (xhigh reasoning), Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, below GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8-2.4T-A95B at 40. Its 89.2 on MCP-Atlas, 61.8 on JobBench and 26.1 on HLE-Diamond without tools are the highest in the table. It trails Kimi K3 on OSWorld Verified (80.4 vs 84.8), scores 76.4 on Terminal-Bench 2.1 versus 85.0 for Kimi K3 and 83.9 for GLM 5.3, and gets 89.0 on GPQA Diamond versus 93.5 for Qwen and Kimi. On safety, Meta reports a StrongREJECT v2 adaptive-jailbreak score of 0.5 (lower is better) against 62.9 to 86.4 for the comparison models, and says the model refused 99.2% of jailbreaks on its internal cyber benchmark. All figures are vendor-reported.
Sources
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta — Hugging Face
- meta-models/Muse-Spark-Open-816B-A42B — Meta (Hugging Face)