An estimated 20 to 30% off the GPU bill for Muse Spark Open.
Muse is built for long-context coding agents, where the KV cache is the main cost of serving. Every product that shrinks the cache raises throughput and cache hit rate together. This page is a preview: the figures are a rough model, and none of our products have been run on Muse yet.
Preview estimate. A rough model of the GPU cost to serve 50 billion tokens a day, with 20 to 30% saved. None of our products have been run on Muse yet.
50 billion tokens a day is about $34k a day.
The input and output split is an assumption: coding-agent traffic is mostly input, so the model uses 95% input and 5% output. Prices are the token-weighted prices across providers, with 68% of input served from cache.
| Tokens per day | Price per million | Revenue per day | |
|---|---|---|---|
| Input, 68% cache hits | 47.5B | about $0.50 | $23.8k |
| Output | 2.5B | $4.25 | $10.6k |
| Total | 50B | $34k, about $12.6M a year |
About $3M to $6M a year of GPUs.
The model assumes 8×B200 nodes, about 800 GB of KV room per node, and an average context near 100k.
Decode
2.5B output tokens a day is about 29k tokens per second. A node holds about 220 sessions of 100k context, and reading the KV takes more than half of each decode step. That is about 3 nodes.
Prefill
32% of input misses the cache, about 15B tokens a day. That is about 1 to 2 nodes.
Total
About 4 to 5 nodes at full load. Real traffic is peaky, so at 30 to 40% utilisation it is about 10 to 15 nodes. At $3 to $4 per GPU-hour that is about $9k to $16k a day.
Where the $0.7M to $1.6M comes from.
| Product | What it changes on Muse | GPU cost cut | Per year |
|---|---|---|---|
| KV cache compression | A smaller KV cache, so more sessions per node and less cache read per decode step | about 15% of decode, if quality holds | $0.3M to $0.6M |
| The same compression on the prefix cache | More cached prefixes fit in memory, so the 68% hit rate could rise to about 75% and prefill work falls about 20% | about 20% of prefill | $0.2M to $0.4M |
| Sparse prefill (1.40× measured at 128k) | Faster long-context prefill | about 25% of prefill | $0.2M to $0.4M |
| KV Exact (lossless), instead of the compressed KV cache | A smaller gain with no accuracy risk | about 8% of decode | $0.15M to $0.3M |
| MoE tiered experts | The weights are already compressed, so the gain is unknown | not counted | none |
| Combined: KV compression, prefix cache and sparse prefill | about 20 to 30% of GPU cost | about $0.7M to $1.6M |
If quality holds with more aggressive KV compression on Muse, the decode saving roughly doubles. Every figure scales with volume: at 500B tokens a day it is 10 times larger.
Preview: in testing.
The costs on this page are estimated, and none of our products have been run on Muse Spark Open yet. The only figure here measured on a real model is the 1.40× sparse prefill speed-up at 128k, and that was not measured on Muse. Results will be published here once they are measured and re-checked.
Serving Muse? Test it with us.
Free benchmark on your own Muse traffic. We measure cache size, quality and capacity per GPU on your prompts, then turn the estimate on this page into one figure for you.
Muse is a trademark of its owner. weights.zip is not affiliated with or endorsed by Meta. All dollar figures are estimates from a rough model.
Free benchmark on your models and traffic.
- Tell us your stackThe models you serve, your context lengths and your serving engine.
- Get your numbersMemory saved, quality against your current format and capacity per GPU, measured on your prompts.
- PilotRun on one workload for 30 days. Pricing is a share of the measured saving.
Built for the models you serve.
Model names and logos are trademarks of their owners and are used only to identify the models tested. weights.zip is not affiliated with or endorsed by DeepSeek, NVIDIA, Z.ai, Meta or MiniMax.