DeepSeek Harness Guide
Back to blog

DeepSeek V4-Flash-Vision-Exp Lands in dsh v0.1.1

DeepSeek shipped a vision model priced identically to V4-Flash, and dsh v0.1.1-rc.1 both adapts it and fixes a Bubblewrap sandbox escape.

Aug 21, 2026DSH Field KitDSH Field Kit

On August 21 DeepSeek shipped two things hours apart, and they are one story: a vision model on the API, and the dsh release that can drive it.

Upgrade first: a sandbox escape is fixed

v0.1.1-rc.1 fixes a Bubblewrap sandbox escape — confined processes could get around their restrictions through /proc/<pid>/root. If you run dsh with the Bubblewrap sandbox wrapped around code you do not trust, that is the reason to upgrade now rather than at your leisure.

It is also the first version-number bump since launch: every release before this one was an 0.1.0 release candidate, and this is 0.1.1-rc.1.

The vision model

deepseek-v4-flash-vision-exp. Experimental, 1M context.

  • Text: on par with V4-Flash, per DeepSeek.
  • Vision: substantially ahead of the text-only Flash on visual agent benchmarks — DeepSeek's own wording puts it near Opus-4.8.
  • Price: identical to V4-Flash. Vision carries no premium.

| per 1M tokens | peak | off-peak | | --- | --- | --- | | Input, cache hit | $0.014 | $0.007 | | Input, cache miss | $0.44 | $0.22 | | Output | $1.32 | $0.66 |

Vendor-reported benchmarks: Terminal Bench 2.1 83.9 (V4-Flash scored 82.7), DeepSWE 59.3, DSBench-Hard 63.6, Chartography 64.3, ZeroBench (Pass@5) 35.0, Agents' Last Exam 27.3.

Two caveats worth keeping. These are DeepSeek's own numbers, unreproduced by anyone else. And Exp means experimental — no stability promise, and experimental endpoints are the ones that get withdrawn.

Using it from dsh

v0.1.1-rc.1's DeepSeek adapter carries the model, so it shows up as a model choice once you upgrade. The plumbing was already there: rc.8 added native image requests, image input for /goal and /plan, and file and session references in the @ menu. The model arrives into a UI that can already feed it pictures.

The rest of v0.1.1-rc.1

  • Fixed a layout glitch when editing text before an @ reference in the composer.
  • Responsive Markdown tables in conversations; honest precision for cache-hit ratios in the 99.x% range; subagent conversation header navigation.
  • ask_user_question answers accept multiple lines, with wrapping and Shift+Enter for newlines.

Release cadence since launch is holding at roughly one RC every two days: rc.7 on the 17th, rc.8 on the 19th, this one on the 21st.

Sources