DeepSeek Adds Vision to V4 Flash for Free: V4-Flash-Vision-Exp Processes Images and Screenshots
DeepSeek released DeepSeek-V4-Flash-Vision-Exp on Friday, an experimental multimodal extension of its V4-Flash model. The variant adds visual input — images, screenshots, diagrams — while maintaining text performance parity with the standard V4-Flash.
V4-Flash’s baseline is established: 79.0% SWE-bench Verified, 82.7% on Terminal-Bench 2.1, NL2Repo 54.2%, Cybergym 76.7%. The vision extension inherits that profile and extends it to multimodal prompts without degrading text task performance, per DeepSeek’s internal testing.
Pricing
Free on DeepSeek’s API during the experimental window, under the same free tier that applies to V4-Flash. No API key changes required; access via model ID deepseek-v4-flash-vision-exp.
Competitive Position
The release targets the mid-tier multimodal market. V4-Flash sits at $0.07/M input, $0.28/M output — significantly below Anthropic’s vision-capable tiers. A free experimental vision model at V4-Flash capability level applies direct pricing pressure on the $1–5/M input range where most enterprise multimodal API budgets sit.
The framing against Anthropic’s Claude Opus 4.8 is explicit in coverage. Opus 4.8 holds 88.6% SWE-bench Verified and supports vision; V4-Flash-Vision-Exp is cheaper at text parity. The trade-off is clear.
What “Exp” Signals
DeepSeek’s experimental release pattern: gather real-world usage before promoting to production. Previous Exp releases (Janus-Pro, V4 Pro during its April beta phase) graduated to full release within 4–8 weeks. The vision capability appears likely to ship in a standard V4-Flash update once the experimental data confirms stability.
V4 Family Current Status
| Model | SWE-bench Verified | Notes |
|---|---|---|
| V4 Pro High | 96.4% | #2 overall on vals.ai |
| V4 Flash | 79.0% | 82.7% Terminal-Bench 2.1 |
| V4-Flash-Vision-Exp | V4-Flash baseline + vision | Free, experimental |
V4 Pro High added to Arena Agent and Text leaderboards on August 19. V4-Flash-Vision-Exp is not yet on any benchmark leaderboard — it will accumulate scores as the experimental window proceeds.