Qwen 3.5 397B-A17BNVIDIA Blackwell
Running Qwen3.5 on GB300 NVL72
Quick answer
Qwen3.5 runs on GB300 NVL72: 67 benchmarked configs so far. See the interactivity ladder below for measured operating points.
Benchmarked configs
67
Serving engines
Dynamo SGLang, Dynamo TRTLLM
Precisions
fp4, fp8
Run dates
2026-07-13 → 2026-08-25
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 | - | - | Dynamo SGLang | fp8 |
| 50 tok/s | - | - | Dynamo SGLang | fp8 |
| 75 tok/s | - | - | Dynamo SGLang | fp8 |
| 100 tok/s | 4,920 | $0.13 | Dynamo SGLang | fp8 |
| 150 tok/s | 1,878 | $0.34 | Dynamo SGLang | fp8 |
| 200 tok/s | 1,034 | $0.62 | Dynamo SGLang | fp8 |
Frequently asked questions
- How fast is Qwen3.5 on GB300 NVL72?
- The InferenceX fleet has 67 benchmarked configs for this pairing; see the interactivity ladder above for the operating points reached so far.
- How much does it cost to serve Qwen3.5 on GB300 NVL72?
- Cost per million tokens is derived from measured throughput and $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model; it appears once this pairing reaches the primary interactivity tier.
- Which serving engines run Qwen3.5 on GB300 NVL72?
- The runs behind this page used Dynamo SGLang, Dynamo TRTLLM in FP4, FP8, 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 Qwen3.5 numbers measured?
- Every number is measured on real GB300 NVL72 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 2026-08-25. The same derivation powers the InferenceX overview leaderboard.