Skip to content

Model Card|DeepSeek-V4-Flash-Vision-Exp

Sep 1, 2026 1 min
TL;DR DeepSeek-V4-Flash-Vision-Exp: 284B total / 13B active-parameter MoE, 1M context, 384K max output; priced identically to plain V4-Flash (Input $0.44 / Output $1.32 at peak, half that off-peak); wins 6 of 7 text-agent benchmarks against its own predecessor (DeepSWE hits 59.3, edging past Opus-4.8's 58.0); multimodal-agent scores close in on Opus-4.8 (trails by 2.9 on ApexBench, actually leads on ZeroBench); each image is capped at 384 tokens / roughly 800×800 resolution, trading fine detail for near-zero cost
Table of Contents
  1. Model Information
  2. Highlights
  3. Benchmark Results
  4. Comparison with Predecessor/Competitors
  5. Implications for Agent Development
  6. Today's Insight
  7. References

🌏 中文版

Model Information

FieldValue
Model IDdeepseek-v4-flash-vision-exp (API call string; HuggingFace repo is deepseek-ai/DeepSeek-V4-Flash-Vision-Exp)
VendorDeepSeek
Parameters284B total / 13B active (sparse MoE; same architecture as DeepSeek-V4-Flash-0731, with a vision module bolted on and continued pretraining)
Context Window1,000,000 tokens (max output 384K tokens)
Input pricing (USD/1M tokens)$0.44 (peak, cache miss) / $0.22 (off-peak, cache miss); cache hit drops further to $0.014 / $0.007
Output pricing (USD/1M tokens)$1.32 (peak) / $0.66 (off-peak)
Open sourceYes (MIT License; weights and reference inference code published on HuggingFace)
Release date2026-08-21 (API launch announcement)
Official announcementDeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live
HuggingFacedeepseek-ai/DeepSeek-V4-Flash-Vision-Exp
FamilyDeepSeek V4.x (first multimodal experimental branch of the V4-Flash line)

Highlights

  • Text-agent capability goes up, not down: wins 6 of 7 text-agent benchmarks against the text-only V4-Flash-0731 it forked from — DeepSeek's own harness measured 59.3 on DeepSWE (a code-repair benchmark), edging past Claude Opus 4.8's 58.0
  • Multimodal-agent performance closes in on Opus-4.8: 36.5 on ApexBench (Pass@1), just 2.9 points behind Opus-4.8's 39.4; 35.0 on ZeroBench (Pass@5) actually beats Opus-4.8's 34.0
  • Images are nearly free: each image is capped at 384 tokens for billing, so at the off-peak rate of $0.22/1M an image costs roughly $0.00008 — community estimates put it at about 2,500 images per dollar
  • Zero-friction API compatibility: reuses V4-Flash's existing Chat Completions / Anthropic Messages / Responses endpoints — swap the model string to deepseek-v4-flash-vision-exp and you're done; DeepSeek Harness 0.1.1, released the same day, already supports it out of the box

Benchmark Results

BenchmarkVision-ExpV4-Flash-0731 (text-only predecessor)Opus-4.8 (strongest competitor)
Terminal Bench 2.183.982.785.0
DeepSWE59.354.458.0
Toolathlon-Verified75.970.376.2
ApexBench (Pass@1, multimodal)36.526.2†39.4
Agents' Last Exam (multimodal)27.325.2†25.7
ZeroBench (Pass@5, multimodal)35.034.0

⚠️ All figures above are DeepSeek's own first-party numbers, measured with its own DeepSeek Harness (Minimal Mode, temperature=1.0, top_p=0.95) on launch day, with no independent third-party reproduction yet. For the two rows marked †, V4-Flash-0731 is a text-only model that simply ignores the multimodal elements of the task during evaluation, so those scores aren't a like-for-like comparison.

Comparison with Predecessor/Competitors

The counter-intuitive part isn't that Vision-Exp gained sight — it's that its text ability improved too. Compared with the V4-Flash-0731 build it was forked from, DeepSWE moved up 4.9 points and Toolathlon-Verified moved up 5.6 points. DeepSeek's own announcement undersells this with a single word ("matches"), but the actual numbers show a real text-capability upgrade rode along with the vision addition, not just a bolted-on encoder. On the multimodal side, once you strip out the artificial gap caused by V4-Flash-0731 simply being unable to see, the distance to Opus-4.8 isn't large: 2.9 points behind on ApexBench, 0.7 behind on Chartography, but actually ahead on both Agents' Last Exam and ZeroBench.

The most aggressive move here is pricing: Vision-Exp is billed on exactly the same rate card as plain V4-Flash — DeepSeek didn't charge a single extra cent for vision. By contrast, Claude's and GPT's vision inputs typically stack additional image-token costs on top of text pricing; DeepSeek instead compresses every image down to a 384-token ceiling, trading resolution for near-zero marginal cost.

Implications for Agent Development

Vision-Exp's positioning is clear: this isn't a general-purpose chat-with-images model, it's built for agents that operate on screens and codebases. OpenRouter's traffic data from its first three days live shows the top consumers are coding harnesses — Claude Code, pi, and similar — with almost nobody running it as a plain chatbot, which lines up with the across-the-board gains on text-agent benchmarks.

  • If you're building a coding agent or CLI agent: it wins 6 of 7 text-agent benchmarks against its predecessor, so swapping the model string to Vision-Exp can be a pure upside even if you never send it a single image — DeepSWE even edges past Opus-4.8
  • If you're building an agent that occasionally needs to "glance at a screenshot" (verifying UI rendered correctly, reading a simple chart): the 384-token / roughly 800×800-resolution cap costs next to nothing, which fits "rough judgment" tasks well
  • Not a fit: anything that needs to read fine print — 8pt text on a receipt, dense dashboards, or engineering drawings — the 800×800 downscale ceiling throws away exactly that detail; DeepSeek also states FIM (fill-in-the-middle) isn't supported yet, so if your pipeline depends on FIM for inline code completion, you can't just swap the model string over

Today's Insight

I assumed adding vision capability usually means trading away some text ability. This release argues otherwise — DeepSeek continued pretraining the same MoE backbone with a vision module attached, and text-agent benchmarks broadly improved rather than regressed. It's a reminder that when evaluating a multimodal upgrade, checking "did the multimodal score go up" isn't enough — you also need to check whether the pure-text tasks were helped or hurt, because the two don't necessarily move together.

References