DeepSeek launches multimodal V4-Flash-Vision-Exp, agent benchmarks close to Opus 4.8
On Aug 21 DeepSeek launched V4-Flash-Vision-Exp, an experimental multimodal vision model priced the same as V4-Flash; text parity with V4-Flash and multimodal agent benchmarks close to Opus 4.8. Files API launched alongside. CMS noted strategic significance as "Android of the agent era."
On August 21 DeepSeek announced via official channels that DeepSeek-V4-Flash-Vision-Exp, a new multimodal vision-understanding model, has launched on the DeepSeek API platform. DeepSeek defines the model as experimental; users must explicitly set the model string to deepseek-v4-flash-vision-exp to access it. Web and App availability were not announced in this release.
On pure text tasks (agent, reasoning, world knowledge) the new model matches DeepSeek-V4-Flash formal release. On agent benchmarks that require vision understanding, multimodal agent capability is close to Anthropic's Claude Opus 4.8. DeepSeek's published comparison shows V4-Flash-Vision-Exp at 83.9 on Terminal Bench 2.1 and 59.3 on DeepSWE, slightly above V4-Flash-0731's 82.7 and 54.4; Chartography 64.3 and ZeroBench Pass@5 35.0, with ApexBench rising from 26.2 to 36.5 and Agents' Last Exam from 25.2 to 27.3.
Pricing matches V4-Flash under the August 16 peak/off-peak schedule: input cache miss $0.22/$0.44, cached input $0.007/$0.014, and output $0.66/$1.32 per million tokens, with peak hours at 01:00–04:00 and 06:00–10:00 UTC. Images convert to tokens by size, with each image capped at 384 tokens.
The API supports Chat Completions, Messages, and Responses invocation formats. Images can be passed by base64 inline, public URL, or via the new Files API. Files API allows uploading an image once and reusing it across requests via file_id. The same-day DeepSeek Harness v0.1.1-rc.1 update ships native image request support for the model adapter and mixed image-text input, while preserving image attachments across MCP and ACP sessions.