Running Kimi K3.1 Pro on GB200 NVL72
Quick answer
Kimi K3.1 Pro sustains 6,994 tokens/s per GPU on GB200 NVL72 at 50 tokens/s per user, which works out to $0.074 per million tokens at hyperscaler pricing, served by Dynamo SGLang. Fastest measured TTFT: 0.6 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
10
Serving engines
Dynamo SGLang
Precisions
fp4
Run dates
2026-11-05 → 2026-11-05
Throughput at every interactivity target
Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on the AgentX agentic coding workload, using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | 9,900 | $0.052 | Dynamo SGLang | fp4 |
| 50 tok/s | 6,994 | $0.074 | Dynamo SGLang | fp4 |
| 75 tok/s | 4,606 | $0.11 | Dynamo SGLang | fp4 |
| 100 tok/s | 3,360 | $0.15 | Dynamo SGLang | fp4 |
| 150 tok/s | 1,981 | $0.26 | Dynamo SGLang | fp4 |
| 200 tok/s | 1,206 | $0.43 | Dynamo SGLang | fp4 |
What serving actually costs
Converting the 50 tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.
| Pricing tier | $ / GPU / hr | $ / 1M tokens |
|---|---|---|
| Owning at Large Hyperscaler Volume | $1.86 | $0.074 |
| Retail | $4.00 | $0.16 |
Frequently asked questions
- How fast is Kimi K3.1 Pro on GB200 NVL72?
- At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, GB200 NVL72 sustains 6,994 tokens/s per GPU serving Kimi K3.1 Pro with Dynamo SGLang in FP4. Peak measured throughput across all configs is 10,567 tokens/s per GPU.
- How much does it cost to serve Kimi K3.1 Pro on GB200 NVL72?
- $0.074 per million total tokens at large-hyperscaler-volume ownership $/GPU/hr pricing, at 50 tokens/s per user. The retail rental tier is tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run Kimi K3.1 Pro on GB200 NVL72?
- The runs behind this page used Dynamo SGLang in FP4, including disaggregated prefill and multi-node serving. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these Kimi K3.1 Pro numbers measured?
- Every number is measured on real GB200 NVL72 hardware by the InferenceX fleet, sweeping concurrency on the AgentX agentic coding workload to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-11-05. The same derivation powers the InferenceX overview leaderboard.