All model and GPU pairings
Qwen 3.5 397B-A17BNVIDIA Blackwell

Running Qwen3.5 on B300

Quick answer

Qwen3.5 sustains 17,277 tokens/s per GPU on B300 at 50 tokens/s per user, which works out to $0.036 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.1 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).

Benchmarked configs

178

Serving engines

SGLang

Precisions

bf16, fp4, fp8

Run dates

2026-05-182026-09-10

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 targetTokens/s per GPU$ / 1M tokensEnginePrecision
30 tok/s9,053$0.069SGLangfp8
50 tok/s17,277$0.036SGLangfp4
75 tok/s14,033$0.045SGLangfp4
100 tok/s11,344$0.055SGLangfp4
150 tok/s8,372$0.075SGLangfp4
200 tok/s5,941$0.11SGLangfp4

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$2.26$0.036
Retail$4.25$0.068

Frequently asked questions

How fast is Qwen3.5 on B300?
At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), B300 sustains 17,277 tokens/s per GPU serving Qwen3.5 with SGLang in FP4. Peak measured throughput across all configs is 97,865 tokens/s per GPU.
How much does it cost to serve Qwen3.5 on B300?
$0.036 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 Qwen3.5 on B300?
The runs behind this page used SGLang 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 B300 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 2026-09-10. The same derivation powers the InferenceX overview leaderboard.

Explore the data