Meta Muse Spark Open: an 816B model for local use, and what the hardware really takes

Meta released the model Muse Spark Open on 15 October 2026. It has 816 billion parameters with about 42 billion active per token and a context window of 1 million tokens, and it handles text, images, video and audio while outputting text. Meta calls it the most capable open model from the US so far. The numbers in this post come from heise's report of 16 October and the MacRumors piece.

Model and hardware figures come from Meta, Apple and heise. The benchmark values were collected by Artificial Analysis. We have not run the model ourselves. (Quelle: Simple Icons, file meta.svg (CC0))

Size and memory

According to heise, the 16-bit version is about 1.6 TB and the partly 4-bit quantized one about 530 GB. MacRumors writes that two Mac Studios with M5 Ultra and 512 GB of unified memory each hold the model together, using Apple's MLX framework. According to heise, MLX distributes the computation across both machines over a Thunderbolt 5 connection. Per heise, Apple measured about 25 tokens per second at an 8,192-token context; that is a vendor figure.

Weight sizes per heise. The memory values are sums of device memory, before context and KV cache. The 512 GB Mac Studios are not expected before the end of October, per heise and MacRumors. (Quelle: heise, 16 Oct 2026)

For our hardware the math is short. Four RTX 5090 with 32 GB each make 128 GB, and the 530 GB version is a little over four times that. With offloading to system RAM it could probably be loaded, but it would not be fast (our own estimate, not measured). Our lane runs smaller models, and Muse Spark Open stays on the watch list until smaller variants appear. Heise puts the two 512 GB machines together at several tens of thousands of euros; MacRumors says the 256 GB Mac Studio already costs over 10,000 US dollars.

License

The Muse Model License allows commercial use in general. It requires the notice "Built with Muse Spark Open" and compliance with a use policy. If you offer the model as an API service, you need a separate agreement with Meta once gross revenue exceeds 20 million US dollars within twelve months, and revenue of affiliated companies counts. MacRumors calls it "limited commercial use". For internal use in a company below that threshold the clause changes nothing; for reselling as a service it does.

How good is it

Artificial Analysis gives the model 39 points on its Intelligence Index, the highest for a US open-weight model. Nvidia's Nemotron 3.5 Super VL (120B) scores 26. Chinese models lead: MiMo-V2.6-Pro with 46 and GLM-5.3 with 45.

Points per heise, citing Artificial Analysis. The six-point gap to GLM-5.3 comes from the agent tests, per heise. (Quelle: heise, 16 Oct 2026)

On agent tasks the model is weak: among the nine top open-weight models it is last in the overall score of four agent tests and fourth in the other six. On AA-Omniscience it reaches 25 points, the highest among open-weight models, with a hallucination rate of 24 percent on incorrect answers, the lowest among open-weight models with at least 30 index points. On average it produces about 40,000 tokens per index task, fewer than any of the eight other leading open-weight models. Less output means lower cost when you pay per token.

API costs

At launch, Together AI, Fireworks and Baseten offer the model at 1.25 US dollars per million input tokens and 4.25 US dollars per million output tokens. Per heise, a benchmark task costs 0.84 US dollars on average, against 2.01 for GLM-5.3 and 0.27 for DeepSeek V4.1 Flash. MacRumors also names Groq, Modal, Nebius and OpenRouter as API platforms, and the weights are on Hugging Face among others. For us, WinLine and banking data go only to local models anyway, so a hosted API is not an option for that work.