Meta's Muse Spark Open (816B-A42B): what it takes to run it locally, from a 2-bit Mac Studio quant to NVFP4 on Blackwell

Meta released Muse Spark Open on October 15, 2026 and called it the most capable open-weight model in the United States. (TechCrunch) It is the full-size sibling of the 30B dense Muse Glimmer and shares its tokenizer and vocabulary. The question builders ask first is simple: what hardware do I need to run it? The answer depends on which format and runtime you pick, so here is what the official Hugging Face launch post says for each.

The model in one table

Spec	Value
Total / active parameters	~816B / ~42B per token
Experts	256 per MoE layer, 8 routed per token
Inputs / output	Text, images, audio, video / text
Context	Up to 1M tokens (max_position_embeddings 1,048,576)
Checkpoints	BF16 and NVFP4

Meta says it trained the model for agentic coding — reading a codebase, writing and debugging code, calling tools — and paired it with its Muse Code terminal agent in demos. (TechCrunch)

The key MoE trade-off: only 8 experts run per token, so it computes like a ~42B model, but all 256 experts must stay in memory, so you pay for 816B parameters of storage.

Memory by format, as published

• BF16: about 1.5 TB of safetensors. In Transformers it was run across 4 nodes of 8 H200 GPUs (TP=32).
• NVFP4: 4-bit, built for NVIDIA Blackwell; runs on a single B200 node. The SGLang reference config uses four GB300 GPUs; the vLLM reference plugin also needs Blackwell and serves a 128K context with --tensor-parallel-size 4.
• llama.cpp GGUF: Hugging Face tested Q2_K, Q4_K_S and Q8 on Apple Silicon. An imatrix 2-bit quant used 304 GB on a 512 GB M3 Ultra and ran at 18 tokens/s.
• MLX (oMLX / mlx-lm): the oQ3e 3-bit quant took about 344 GB on the same M3 Ultra, with ~225 tok/s prefill and ~21 tok/s generation (1,024-token prompt, 128 generated tokens). Hugging Face notes even 3-bit needs more than 256 GB of RAM.

So “runs on Macs” means a 512 GB-class Mac Studio, not a laptop. Reporting differs here: TechCrunch, citing Engadget, says local use requires at least 1 TB of RAM. That fits the BF16 weights. The Hugging Face numbers above are for 2–3-bit quants. If you have less than 256 GB of unified memory, Muse Glimmer 30B is still the Muse model to run.

KV cache at long context

Grouped-query attention (16 KV heads of 64 dims, 62 layers) keeps the cache at about 250 KB per token in BF16. That is roughly 266 GB for a full 1M-token context, on top of the weights. That is why the vLLM recipe caps serving at 128K. If you run long agent sessions in llama-server, the same cache-reuse issues covered in our KV cache reuse guide apply.

Getting started commands (from the launch post)

llama.cpp, with the quant as a suffix:

llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K

mlx-lm, using the PR build referenced in the post:

mlx_lm.generate --model mlx-community/Muse-Spark-Open-816B-A42B-oQ3e \
  --prompt "Explain convolutional networks like I am 5" --max-tokens 5000

Hugging Face Inference Endpoints also has a validated NVFP4 SGLang recipe: 8× RTX Pro 6000 (98 GB) for 131,072 context, or 8× H200 for the full 1,048,576.

Two behaviours to know before you wire it into an agent

1. Reasoning cannot be turned off. The levels are low/medium/high/xhigh, mapped to 16/64/256/512, and the default is xhigh. Set “low” for easy tasks to save tokens.
2. It uses a Harmony-style, channel-based chat template (first used by gpt-oss). The self channel carries reasoning, commentary carries text the user sees around tool calls, and user carries the final answer. If output quality looks off, check which channel the model routed to.

Bottom line

Muse Spark Open is the first Muse model that needs serious hardware. For Apple Silicon, plan on 512 GB and a 2–3-bit quant at around 18–21 tok/s. For NVIDIA, the NVFP4 checkpoint on Blackwell is the reference path. Everything else in the family stays in Glimmer territory.

Sources: Hugging Face launch post · TechCrunch · 9to5Mac