The fastest way to serve DeepSeek V4 Flash

DeepSeek released V4 Flash 0731 on July 31. Two days later we published a validated serving package for it. On the same four B200 GPUs, in the same measurement session, the v4turbo package serves 362.86 output tokens per second to a single user. A stock deployment serves 113.37.

That is the whole announcement. The rest of this post is the evidence: the measured pair, the serving curve including the point where the package loses, what the market serves today, what the speed is for, and what the package costs.

The claim, with its conditions

No provider benchmarked on the public leaderboard for this model serves it faster today, and that sentence needs its conditions attached. Artificial Analysis lists no provider speed benchmarks at all for the current 0731 release; its provider page says benchmarks are not available and names DeepSeek's own API as the sole provider (accessed 2026-08-02). On the preview-era board for the non-reasoning variant, the fastest listed provider is Makora at 241.4 output tokens per second. Our measured number is 362.86, and it is a single-stream measurement on an idle dedicated node with the exact request profile below, while leaderboard providers are measured under production load. Fastest, with the receipts and the caveat in the same breath.

The protocol behind our number: 1024 input tokens, 1024 output tokens, temperature 0, streaming, 8 timed requests per arm, both arms back to back in one window in one container on the same four B200 cards, request profiles verified identical before any number rendered. One arm is a stock deployment of the same model on the same node. The other is the v4turbo package. The only changed variable is the serving configuration, and the raw artifacts are committed.

The headline pair

The serving curve, including where we lose

Single-stream speed is one operating point. The serving curve is the rest. We measured 32768-token prompts with 2048-token outputs at concurrency 1, 8, and 32, per-GPU output tokens per second, both arms.

The recipe is a curve and a rule, not a single setting. The accelerated configuration is faster at concurrency 1 and 8. At concurrency 32 the standard configuration holds the higher aggregate, 397.74 per GPU against 370.43, so the rule runs it there and the node keeps 1590.94 output tokens per second. You are never on the losing side of this chart if you follow the rule it ships with.

What the market serves today

The claim section above carries the conditions; this chart is the picture. The three provider bars are what Artificial Analysis lists for the preview-era variant of this model. The highlighted bar is our measured single-stream number.

Where the speed matters

• Coding agents and tool-use loops. Agent steps are serial, each one waits for the last, so per-user tokens per second sets wall-clock time. A 1024-token step streams in about 3.1 seconds instead of 9.2.
• Interactive assistants. A person watching tokens stream reads the difference between 113 and 363 tokens per second directly, and it costs about 13 ms at the first token.
• High-volume batch. The operating rule runs the standard configuration here, which holds 397.74 output tokens per second per GPU at concurrency 32 (the accelerated configuration measures 370.43 at that point). Both configurations and the measured curve ship in the package, so batch traffic keeps the highest aggregate. Measure your own operating point before you commit.

Own your AI

Every closed API in the table below is a rented dependency: rented latency, rented rate limits, rented deprecation schedules. The table is what list prices look like beside a model you can own.

Serving it yourself means renting the node instead. The measured target for this package is four B200 GPUs, and three providers publish on-demand rates for exactly that.

One piece of arithmetic, with its basis stated. At the measured concurrency 32 operating point one four-GPU node aggregates 1590.94 output tokens per second in the stock configuration and 1481.71 with the package. Batch is where stock wins, so the batch arithmetic uses the stock number: about 5.7 million output tokens per hour. Divide the node rent by that and a fully busy node lands between $4.11 and $4.99 per 1M output tokens, and each of those requests also carried a 32768-token prompt the same dollars paid to process. The assumptions are visible: the node stays busy, and your traffic resembles that workload. Redo this arithmetic against your own measured curve before you commit.

DeepSeek's own API lists $0.28 per 1M output tokens, far below that node arithmetic, so the case for self-hosting this model is not raw price per token. It is the list that follows.

• Your prompts and outputs stay on hardware you control.
• There are no rate limits except the ones you set.
• There are no model deprecations; the weights you validated today are the weights you serve next year.
• Latency is yours to control, down to where the node physically sits.
• The model is MIT licensed, so it is yours to keep. We read the license from the upstream repository on 2026-08-01.

What you get

The v4turbo package is the validated serving configuration for DeepSeek V4 Flash 0731 on a four B200 node, sold as a finished artifact.

• The validated serving configuration, exactly as measured.
• The measured proof: the headline pair, the serving curve, and committed raw artifacts behind every number in this post.
• Deploy files for the measured 4x B200 target.
• Reproduction instructions, so you can re-run the measurement on your own node before you trust ours.

The measured serving results and reproduction context are documented in this article.

Sources

• Measured performance numbers: RunInfra measurement session of 2026-08-01, one 4x B200 node, with the measurement conditions documented in this article.
• Artificial Analysis, deepseek-v4-flash provider board (no provider speed benchmarks; sole listed provider DeepSeek at $0.14 in / $0.28 out): https://artificialanalysis.ai/models/deepseek-v4-flash/providers, accessed 2026-08-02.
• Artificial Analysis, deepseek-v4-flash-non-reasoning provider board (Makora 241.4 tok/s, DeepSeek 107.2 tok/s, CoreWeave 40.8 tok/s): https://artificialanalysis.ai/models/deepseek-v4-flash-non-reasoning/providers, accessed 2026-08-02.
• Anthropic pricing (Claude Fable 5 $10.00 in / $50.00 out, Claude Opus 5 $5.00 / $25.00, Claude Sonnet 5 $2.00 / $10.00 introductory through 2026-08-31, then $3.00 / $15.00): https://platform.claude.com/docs/en/about-claude/pricing, accessed 2026-08-02.
• OpenAI pricing, Standard tier (gpt-5.6-sol $5.00 in / $30.00 out, gpt-5.6-terra $2.00 / $12.00, gpt-5.6-luna $0.20 / $1.20): https://developers.openai.com/api/docs/pricing, accessed 2026-08-02.
• Google pricing (Gemini 3.1 Pro Preview $2.00 in / $12.00 out for prompts up to 200k tokens, $4.00 / $18.00 above): https://ai.google.dev/gemini-api/docs/pricing, accessed 2026-08-02.
• DeepSeek API pricing (deepseek-v4-flash $0.14 in on cache miss, $0.0028 on cache hit, $0.28 out; 2x peak-hour pricing announced, effective date pending): https://api-docs.deepseek.com/quick_start/pricing, accessed 2026-08-02.
• 4x B200 on-demand rental: RunPod $5.89 per GPU-hour (https://www.runpod.io/pricing), Lambda $6.79 per GPU-hour (https://lambda.ai/service/gpu-cloud/pricing), Nebius $7.15 per GPU-hour (https://nebius.com/prices), all accessed 2026-08-02.
• DeepSeek-V4-Flash-0731 weights and MIT license: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731, license read 2026-08-01.