Running Qwen3.8-Flash-Next on GB300 NVL72
Quick answer
Qwen3.8-Flash-Next sustains 117,140 tokens/s per GPU on GB300 NVL72 at 50 tokens/s per user, which works out to $0.005 per million tokens at hyperscaler pricing, served by SGLang. Fastest measured TTFT: 0.5 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
8
Serving engines
SGLang
Precisions
fp4
Run dates
2026-10-09 β 2026-10-09
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 | 138,311 | $0.005 | SGLang | fp4 |
| 50 tok/s | 117,140 | $0.005 | SGLang | fp4 |
| 75 tok/s | 87,111 | $0.007 | SGLang | fp4 |
| 100 tok/s | 66,162 | $0.010 | SGLang | fp4 |
| 150 tok/s | 32,353 | $0.020 | SGLang | fp4 |
| 200 tok/s | 17,303 | $0.037 | 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 | $2.31 | $0.005 |
| Retail | $5.00 | $0.012 |
Frequently asked questions
- How fast is Qwen3.8-Flash-Next on GB300 NVL72?
- At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, GB300 NVL72 sustains 117,140 tokens/s per GPU serving Qwen3.8-Flash-Next with SGLang in FP4. Peak measured throughput across all configs is 141,337 tokens/s per GPU.
- How much does it cost to serve Qwen3.8-Flash-Next on GB300 NVL72?
- $0.005 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 Qwen3.8-Flash-Next on GB300 NVL72?
- The runs behind this page used SGLang in FP4. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these Qwen3.8-Flash-Next numbers measured?
- Every number is measured on real GB300 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 2026-10-09. The same derivation powers the InferenceX overview leaderboard.