Muse Spark Open VRAM requirements at Q4 and 1M context

How this is calculated

The config has 62 layers with hidden size 8,192, 16 KV heads at head dim 64, and 256 experts with 8 used per token. The layer mix is what shapes the cache. Of the 62 layers, 46 use a 2,048 token sliding window and 16 use full attention (layers 1, 5, and so on up to 61). A sliding window layer never keeps more than its window, so those 46 layers stop growing at about 0.39 GB of cache together. Only the 16 full layers grow with the prompt. At 1,048,576 tokens they add up to 69.1 GB of FP16 cache, and at 32K the whole cache is 2.53 GB. The drop from 1M to 32K saves 73 GB of the total. That's a real saving, since the cache is 12% of the estimate at 1M. A supported 8-bit cache would bring the 1M total down to about 541 GB, but that depends on your engine. Serving several users changes the math too. Each sequence needs its own cache while the weights are shared, so four sequences at 128K tokens come to 542 GB. The weights stay the largest line at 79% of the budget, so a shorter context helps the cache and nothing else. The estimate follows the published config. It leaves out vision, video and audio processing, the engine's prefill workspace and any speculative decoding buffers.

Verdict

A single node comes close here. Eight cards of 80 GB give 640 GB, which is above the full context estimate of 579 GB on paper. The total takes 8 cards of 80 GB, and a 32K session takes 7. On paper isn't the same as working. Sharding is rarely perfectly even, and prompt processing needs temporary memory on top of the cache, so leave margin and test at your real context length. Cards of 192 GB bring the count down to 4 for the full context. The weights are gated on Hugging Face under a custom license, so read it and accept it before you download anything. Then compare the real file sizes of the checkpoint you choose with the 457 GB assumed here, since Q4_K_M is a sizing factor and not a promise about any one file.