# DeepSeek V4 Flash 0731 — Sovereign Coding Room

Status: **BENCHMARKED / CAPACITY REVIEW**. This is not a live commercial SKU.

The official 304B mixed-FP4/FP8 checkpoint ran on one NVIDIA B300 288 GB with
vLLM 0.25.0 and DSpark. The best measured configuration used eight concurrent
requests. It reached 169.74 aggregate output tok/s with 6.03-second median
request latency. Sixteen concurrent requests regressed to 104.24 tok/s and
19.64-second latency.

Coding smoke: 4/5 first-pass tasks passed; the failed async task also failed one
traceback-guided repair. Every generated patch therefore requires compilation
and deterministic tests.

## Use boundary

Use for confidential repositories, sovereign demonstrations, and bounded
research batches. Do not use for heartbeats, cron, routine chat, or unattended
production merges. Queue at least four tasks, batch at eight, allow one repair,
and tear down after five idle minutes.

## Start

On a RunPod B300 CUDA 13 / Ubuntu 24.04 base image:

```bash
curl -fsSLO https://genie.locker/recipes/deepseek-v4-flash-0731/serve.sh
chmod 700 serve.sh
./serve.sh
```

Then require `GET http://127.0.0.1:8000/v1/models` to return
`deepseek-v4-flash-0731` before sending work.

## Honest economics

The tested B300 cost $7.39/hour. Best measured eight-way utilization modeled to
$12.09 per million output tokens, versus DeepSeek's published hosted price of
$0.28. The roughly 43x premium buys sovereignty, not cheap tokens.

See `recipe.json` for the complete evidence card and known boundaries.