Muse Spark 1.3/Open 816B
Cheapest GPU for Muse Spark 1.3/Open
GB200 NVL72 is the cheapest at $0.013 per million tokens, 1% below B200.
Measured on the AgentX agentic coding workload at a matched interactivity target of 125 tokens/s per user. Hardware without a measurement at this operating point is not ranked.
| Rank | GPU | Tokens/s per GPU | $ / 1M tokens | Precision | Engine |
|---|---|---|---|---|---|
| 1 | GB200 NVL72 | 38,936 | $0.013 | fp4 | SGLang |
| 2 | B200 | 35,736 | $0.013 | fp4 | SGLang |
| 3 | GB300 NVL72 | 43,643 | $0.015 | fp4 | SGLang |
| 4 | B300 | 39,693 | $0.016 | fp4 | SGLang |
Not measured at 125 tokens/s per user: MI355X.
Optimizing for speed instead of cost? See the fastest GPU for Muse Spark 1.3/Open
Methodology
Every number on this page is a measurement, not a spec-sheet estimate. The InferenceX fleet serves Muse Spark 1.3/Open on real hardware with community serving engines, sweeping concurrency to trace each platform's throughput-versus-interactivity frontier.
Platforms are then read at the same operating point (125 tokens/s per user) so the comparison is iso-interactivity: a GPU cannot win by quoting throughput at an unusably slow per-user speed. Rankings default to 125 tokens/s per user, the speed current runs cover on the most hardware, and use another speed only when it ranks more GPUs for this model. Cost converts measured throughput to $ per million total tokens using $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model.
The derivation is shared with the InferenceX overview leaderboard, and results re-run continuously, so this ranking updates as new engine releases and configs land.
Frequently asked questions
- What is the cheapest GPU to run Muse Spark 1.3/Open?
- As of the latest benchmark runs, GB200 NVL72 is cheapest at $0.013 per million total tokens on the AgentX agentic coding workload, at hyperscaler $/GPU/hr pricing and a matched interactivity target of 125 tokens/s per user.
- How is cost per million tokens calculated?
- Measured throughput per GPU at the 125 tokens/s per user tier is converted to $ per million total (input plus output) tokens using hyperscaler $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model. Slower interactivity targets or cheaper rental tiers change the absolute numbers but rarely the order.
- How often is this ranking updated?
- Benchmarks re-run continuously on the InferenceX cluster fleet; the newest result feeding this ranking landed on 2026-11-11. The page always renders the latest data.