Muse Spark Open Puts Meta’s 816B Coding Model on Two Mac Studios, With a $20M API Tripwire

Meta released the weights for Muse Spark Open on October 15, 2026, an 816-billion-parameter mixture-of-experts model that activates about 42 billion parameters per token and accepts text, images, audio and video. Meta’s research blog calls it “the most capable US open-weight model to date” and pitches it at coding and agentic work. The checkpoints sit on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B in BF16 and a separate NVFP4 build, behind a manual access form. The Muse Model License is not an open-source license: it is Meta’s own custom license, not MIT, Apache or any OSI-approved terms, and it carries a revenue clause aimed at large model API providers.

Inside the 816B mixture

The model card, authored by Meta Superintelligence Labs, lists a 62-layer language model with a hidden size of 8192, 128 query heads and 16 key-value heads in grouped-query attention, and a head dimension of 64. Every layer swaps the dense feed-forward network used in the smaller Muse Glimmer for a block of 256 experts, with a router choosing 8 per token. Hugging Face’s launch write-up says the experts hold roughly 97% of all parameters. That is why the model computes like a 42B system but still needs memory for all 816B weights.

A 1.8B-parameter perception encoder turns images and video frames into tokens, and an audio encoder emits one token every 80 milliseconds, or 12.5 tokens per second of speech. The context window is 1M tokens at inference, shared by text, image, video and audio tokens. Hugging Face estimates a full million-token BF16 KV cache at around 266 GB.

Reasoning cannot be switched off. Developers choose low, medium, high or xhigh effort, and xhigh is the default. The chat template routes internal reasoning, user-visible commentary between tool calls, and final answers through separate channels.

Hardware math, from data center to desk

According to the model card, self-hosting needs at least 700 GB of GPU memory for the NVFP4 checkpoint and 2 TB for BF16. Hugging Face says the BF16 files come to about 1.5 TB and run in Transformers across four nodes of eight H200 GPUs, while NVFP4 fits a single B200 node. Its SGLang reference runs NVFP4 on four GB300 GPUs.

Meta says Muse Spark Open is the first model in its class to run on two Apple M5 Ultra Mac Studios using MLX. Each machine can give up to 512 GB of unified memory to the GPU, so two of them can hold the whole model. Hugging Face’s own tests on an older M3 Ultra with 512 GB show what quantization buys. A 3-bit oMLX quant used about 344 GB and generated roughly 21 tokens per second. An imatrix 2-bit llama.cpp quant used 304 GB and ran at 18 tokens per second.

Meta lists SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth as supported engines. Hosted API access launched with Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Benchmarks that do not sweep the table

Meta’s comparison table puts Muse Spark Open ahead on several agentic and safety rows, not all. All figures are vendor-reported. It scores 89.2 on MCP-Atlas, the top figure in the table, and 26.1 on HLE-Diamond without tools. On the Artificial Analysis Intelligence Index it posts 39, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8-2.4T-A95B at 40. Kimi K3 also leads it on OSWorld Verified (84.8 versus 80.4) and Terminal-Bench 2.1 (85.0 versus 76.4).

The safety rows show the widest gap. On StrongREJECT v2 with adaptive attacks, where lower is better, Meta reports 0.5 for Muse Spark Open against 62.9 to 86.4 for the other models it measured. The card says this open release received extra safety post-training so it holds up as a standalone model without system guardrails.

What the Muse Model License allows

The grant itself is broad: use, modify, fine-tune, distill, merge, redistribute and sell, as long as the Usage Policy travels with the model. There are two notable strings. Products using the model or its derivatives must show “Built with Muse Spark Open” in the interface or documentation. And any company offering models as a service whose group revenue tops $20 million over 12 months must notify Meta and sign a separate agreement before serving Muse Spark Open or derivatives. That clause explicitly exempts internal use, research, and using the model inside a company’s own product that is not a model API. The Usage Policy says the model is not intended for anyone under 18 and lists prohibited uses from weapons development to malware.

For most teams the harder limit is hardware, not paperwork. A 42B-active model still needs 700 GB or more of GPU memory to serve, and the desktop route assumes two of Apple’s largest Mac Studios.