All model and GPU pairings
DeepSeekv4 Pro 0813 1.6TNVIDIA Blackwell

Running DeepSeek V4 Pro on B300

Quick answer

DeepSeek V4 Pro sustains 33,747 tokens/s per GPU on B300 at 50 tokens/s per user, which works out to $0.019 per million tokens at hyperscaler pricing, served by vLLM. Fastest measured TTFT: 0.1 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).

Benchmarked configs

171

Serving engines

Dynamo SGLang, Dynamo vLLM, SGLang, TRTLLM, vLLM

Precisions

fp4

Run dates

2026-06-112026-09-01

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 targetTokens/s per GPU$ / 1M tokensEnginePrecision
30 tok/s66,050$0.010SGLangfp4
50 tok/s33,747$0.019vLLMfp4
75 tok/s12,902$0.049SGLangfp4
100 tok/s10,530$0.060vLLMfp4
150 tok/s5,021$0.13vLLMfp4
200 tok/s--SGLangfp4

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$0.019
Retail$4.25$0.035

Frequently asked questions

How fast is DeepSeek V4 Pro on B300?
At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, B300 sustains 33,747 tokens/s per GPU serving DeepSeek V4 Pro with vLLM in FP4. Peak measured throughput across all configs is 108,881 tokens/s per GPU.
How much does it cost to serve DeepSeek V4 Pro on B300?
$0.019 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 Pro on B300?
The runs behind this page used Dynamo SGLang, Dynamo vLLM, SGLang, TRTLLM, vLLM in FP4, 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 DeepSeek V4 Pro numbers measured?
Every number is measured on real B300 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 2026-09-01. The same derivation powers the InferenceX overview leaderboard.

Explore the data