DeepSeek has quietly shipped an experimental vision variant of its budget flagship: deepseek-v4-flash-vision-exp. The new model accepts image input alongside text, extending the V4-Flash family beyond pure text workloads.
What’s New
According to DeepSeek’s API documentation, the current lineup now includes:
- deepseek-v4-flash — updated to DeepSeek-V4-Flash-0731
- deepseek-v4-pro — updated to DeepSeek-V4-Pro-0813
- deepseek-v4-flash-vision-exp — the new experimental vision model that additionally accepts image input
Calling methods remain unchanged for the updated models — existing deepseek-v4-flash and deepseek-v4-pro names automatically route to the latest versions.
Why Vision on Flash Matters
DeepSeek’s Flash models are the workhorses of budget AI — they top global token usage charts precisely because they combine low cost with strong capability. Adding vision to this tier means:
- Multimodal apps get cheaper — image understanding without paying premium model prices
- Agent capabilities expand — coding and computer-use agents can now process screenshots and visual context on the same cheap backend
- Competitive pressure — the move matches the multimodal default that frontier labs (GPT, Gemini, Claude, Qwen) established long ago
Agent Integration Momentum
The documentation also highlights DeepSeek Harness, now in developer preview for agent harness developers, and notes that the DeepSeek API works directly with tools like Claude Code, GitHub Copilot, and OpenCode — no code changes required. The Anthropic API format is supported natively alongside the OpenAI format.
This positioning reflects DeepSeek’s strategy: be the inexpensive, drop-in backend for the agent ecosystem others are building.
Context
The vision experiment follows a busy stretch for DeepSeek: V4-Flash topped global token usage earlier this summer, and the V4 family’s agent capabilities have made it a favorite for cost-conscious developers. As an “experimental” release, the vision model may change quickly — but it signals that multimodality is no longer a premium feature. It’s table stakes, even at the budget tier.