DeepSeek Harness Minimal Mode: The Two-Tool Benchmark Rig

Minimal mode is the smallest runtime configuration of DeepSeek Harness: a two-tool coding agent built specifically for benchmarking models in a minimal environment^[1][2].

It is the configuration DeepSeek used for every public Code Agent benchmark of V4-Flash-0731 — the scores that shipped with the July 31, 2026 release and the +645% DeepSWE headline^[3].

The entire tool surface is two tools^[1][2]:

That is everything. No web search, no skills, no subagents, no vision, no extra sandboxing layers in the default composition. The agent solves coding tasks with the shell and the editor only^[1][2].

• A persistent bash shell — the agent can run arbitrary commands and keep state between invocations
• str_replace_editor — a minimal file editor (string replacement based)

The point of minimal mode is attribution. When a benchmark number is reported, reviewers need to know how much of it came from the model versus the harness's extra machinery^[2].

A full agent environment (web search, skills, subagents) makes scores hard to interpret: did the model solve the task, or did the harness's tools? Minimal mode reduces the harness to its thinnest possible layer, so the score mostly reflects the model's coding ability under a controlled, reproducible setup^[2][3].

It also levels the playing field for third-party reproduction: install dsh, switch to the minimal profile, point it at any OpenAI-compatible model, and run the same benchmark suite under the same two tools^[2].

The exact settings used for the official runs, from the July 31 changelog^[3]:

The same parameters appear in the Hugging Face model card for V4-Flash-0731. They define the reproducibility contract for the published agent benchmarks^[3].

Because dsh is open source, the rig is now runnable by anyone^[1][2]:

For a fair model-vs-model comparison, keep the harness constant (minimal profile, max effort, topp 0.95, temp 1.0) and vary only the model — the methodology third-party reviewers recommend^[4].

# install (one line)
npx @deepseek-ai/dsh web

# point at the model under test (any OpenAI-compatible provider)
# Settings → Models → add provider

# launch the minimal profile headless on a task
dsh --profile headless "implement the function, write tests, run them"

Where minimal mode fits in the runtime family^[1]:

Choose minimal mode when you want measurement or reproducibility; choose Standard for real work and Code for deterministic automation^[1].