Meta Releases Muse Spark Open With Open Weights
816 billion parameters, 42 billion active and a 1 million token context: Meta's new model runs on two Mac Studios instead of a data center.
In short
Muse Spark Open is Meta's new downloadable frontier model: a multimodal mixture-of-experts network with 816 billion parameters, 42 billion of them active per token, which The Decoder reports runs locally on two Mac Studios holding 512 GB each.
At a glance
- 816 billion total parameters, 42 billion active per token; 62 layers, 256 experts per MoE layer, 8 routed per token.
- Context window up to 1 million tokens; accepts text, image, audio and video input and returns text only.
- BF16 checkpoint is about 1.5 TB; the full 1M-token KV cache alone needs 266 GB. The NVFP4 build fits one B200 node.
- The Decoder reports local use on two Mac Studios with 512 GB each; on an M3 Ultra, 18 tokens/s via llama.cpp at 2-bit.
- Weights are free with a required "Built with Muse Spark Open" label; API resellers above $20 million revenue need their own deal.
Meta released Muse Spark Open on October 15, 2026 — a multimodal mixture-of-experts model with 816 billion parameters, only 42 billion of which fire per token, and with downloadable weights.
The architecture in plain numbers
The Hugging Face model card describes 62 layers at a hidden dimension of 8192. Each MoE layer holds 256 experts and routes 8 of them per token. Attention uses 128 query heads against 16 key-value heads, the grouped-query layout that keeps the cache affordable.
Context runs to 1 million tokens. The model takes text, images, audio and video as input and answers in text only — it is multimodal in, text out.
The hardware story is the news
At BF16 the checkpoint is roughly 1.5 TB of safetensors. Meta's own deployment guidance is four nodes of 8 H200 GPUs for BF16, or a single B200 node for the 4-bit NVFP4 build. The KV cache for a full million tokens takes 266 GB on its own at BF16.
The other end of the range is what drew attention. TechCrunch reports Meta tuned the release for Apple silicon and puts the floor at about 1 terabyte of RAM; The Decoder spells that out as two Mac Studios with 512 GB of unified memory each.
One caveat worth keeping: every throughput figure in Meta's own post comes from an M3 Ultra — 18 tokens per second through llama.cpp at 2-bit, and roughly 225 tokens/s prefill with 21 tokens/s generation through MLX. We have no published numbers for the M5 Ultra that TechCrunch names as the target chip.
Strong on agents, weaker on code
Meta claims category leads in Agentic, Knowledge & Reasoning and Multimodal work, while conceding ground on coding benchmarks. Per The Decoder, it beats Qwen 3.8, GLM 5.3 and Kimi K3 on JobBench, MCP-Atlas, HLE-Diamond and ZeroBench, yet still sits below Meta's paid Muse Wonder Full tier.
Meta told TechCrunch this is the most capable open-weight model in the United States. The qualifier does the real work in that sentence: the open-weight frontier has lately been set in China, and this launch is aimed squarely at it.
Free, with a revenue line
Use is free according to The Decoder, on condition that outputs and products carry a "Built with Muse Spark Open" label. Commercial API providers clearing more than $20 million a year need a separate agreement with Meta.
We could not read the license text itself in the Hugging Face page we fetched, so both the labeling duty and the revenue threshold rest on The Decoder's reporting rather than on a document we verified.
What you can actually do with it
Meta pairs the model with its Muse Code terminal and demonstrates bug fixes, feature work and complete software builds. TechCrunch also describes a run where the model reviewed and cropped footage inside DaVinci Resolve.
Fine-tuning is supported through LoRA in TRL: a sample run of 500 steps on 64 H100 GPUs took about two hours at 8,000 tokens per second. Inference paths exist for Transformers, llama.cpp, SGLang, vLLM and MLX.
Without a data center, the cost shows up as latency. At 18 to 21 tokens per second, this is a model you read along with — not one you hand a long agent loop and walk away from.
FAQ
How many parameters does Muse Spark Open have?
816 billion in total, but as a mixture-of-experts model it activates only 42 billion per token, routing 8 of 256 experts in each MoE layer.
What hardware do you need to run Muse Spark Open locally?
TechCrunch puts the floor near 1 terabyte of RAM, and The Decoder specifies two Mac Studios with 512 GB each. For servers, Meta lists four nodes of 8 H200 GPUs at BF16 or a single B200 node for NVFP4.
Can you use Muse Spark Open commercially?
The Decoder says yes and for free, provided you carry the "Built with Muse Spark Open" label. API providers above $20 million in annual revenue need a separate Meta agreement. We did not verify the license text directly.
