Running DeepSeek V4.1 Flash on H100
Quick answer
DeepSeek V4.1 Flash sustains 1,995 tokens/s per GPU on H100 at 50 tokens/s per user, which works out to $0.16 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.3 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
18
Serving engines
SGLang, vLLM
Precisions
fp4
Run dates
2026-09-18 β 2026-09-24
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 | 2,694 | $0.12 | SGLang | fp4 |
| 50 tok/s | 1,995 | $0.16 | SGLang | fp4 |
| 75 tok/s | 11,385 | $0.029 | vLLM | fp4 |
| 100 tok/s | 9,711 | $0.033 | vLLM | fp4 |
| 150 tok/s | 7,153 | $0.045 | vLLM | fp4 |
| 200 tok/s | 5,126 | $0.063 | vLLM | 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 | $1.17 | $0.16 |
| Retail | $2.00 | $0.28 |
Frequently asked questions
- How fast is DeepSeek V4.1 Flash on H100?
- At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, H100 sustains 1,995 tokens/s per GPU serving DeepSeek V4.1 Flash with SGLang in FP4. Peak measured throughput across all configs is 12,673 tokens/s per GPU.
- How much does it cost to serve DeepSeek V4.1 Flash on H100?
- $0.16 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 DeepSeek V4.1 Flash on H100?
- The runs behind this page used SGLang, vLLM in FP4. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these DeepSeek V4.1 Flash numbers measured?
- Every number is measured on real H100 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 2026-09-24. The same derivation powers the InferenceX overview leaderboard.