Running Qwen3.5 on B200
Quick answer
Qwen3.5 sustains 12,503 tokens/s per GPU on B200 at 50 tokens/s per user, which works out to $0.038 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.2 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
199
Serving engines
SGLang, TRTLLM
Precisions
bf16, fp4, fp8
Run dates
2026-02-20 → 2026-08-13
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 a single-turn chat workload (8k input / 1k output), using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | 8,501 | $0.057 | SGLang | fp8 |
| 50 tok/s | 12,503 | $0.038 | SGLang | fp4 |
| 75 tok/s | 9,928 | $0.048 | SGLang | fp4 |
| 100 tok/s | 7,924 | $0.061 | SGLang | fp4 |
| 150 tok/s | 5,679 | $0.085 | SGLang | fp4 |
| 200 tok/s | 4,360 | $0.11 | 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 |
|---|---|---|
| Hyperscaler | $1.73 | $0.038 |
| Neocloud | $2.07 | $0.046 |
| Retail | $2.60 | $0.058 |
Frequently asked questions
- How fast is Qwen3.5 on B200?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), B200 sustains 12,503 tokens/s per GPU serving Qwen3.5 with SGLang in FP4. Peak measured throughput across all configs is 71,322 tokens/s per GPU.
- How much does it cost to serve Qwen3.5 on B200?
- $0.038 per million total tokens at hyperscaler $/GPU/hr pricing, at 50 tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run Qwen3.5 on B200?
- The runs behind this page used SGLang, TRTLLM in BF16, FP4, FP8. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these Qwen3.5 numbers measured?
- Every number is measured on real B200 hardware by the InferenceX fleet, sweeping concurrency on a single-turn chat workload (8k input / 1k output) to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-08-13. The same derivation powers the InferenceX overview leaderboard.