DeepSeek-V4-Flash-Vision-Exp

// Overview

DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal member of DeepSeek's V4 series, announced on 21 August 2026 and served from the DeepSeek API Platform under the model id `deepseek-v4-flash-vision-exp`. It is the first V4 model that accepts images: a request can mix text and image content, and DeepSeek reports that on text-only work — agents, reasoning and world knowledge — it matches DeepSeek-V4-Flash.

Images can be supplied three ways: inline as base64 data, as an external HTTP(S) URL, or as a reference to a file uploaded through the Files API that shipped alongside the model (uploading is free, and one uploaded image can be reused by file id across many requests). The API accepts JPEG, PNG, GIF and WebP, detected from file content rather than the extension, and allows up to 600 images per request. Images are only valid in `user` messages; a `detail` parameter (`low`, `high`, `original`, `auto`) controls how aggressively each one is processed, with `low` downscaling to 512×512.

For billing and context accounting each image is resized to roughly 800×800 with its aspect ratio preserved and costs at most 384 tokens, charged at the ordinary V4-Flash token rate. The model keeps the V4 line's 1M-token context and 384K-token maximum output, is reachable through the Chat Completions, Messages and Responses APIs, and got same-day support in DeepSeek Harness 0.1.1. Unlike DeepSeek-V4-Flash and DeepSeek-V4-Pro, no weights have been published for this checkpoint — it is API-only and explicitly labelled experimental.

// Benchmarks

This model's scores

1. Terminal Bench 2.183.9%
2. Toolathlon-Verified75.9%
3. Cybergym75.3%
4. DSBench-Hard63.6%
5. DeepSWE59.3%
6. NL2Repo57.7%
7. Chartography (multimodal)64.3%
8. ApexBench Pass@1 (multimodal)36.5%
9. ZeroBench Pass@5 (multimodal)35%
10. Agents' Last Exam (multimodal)27.3%
11. AutomationBench (public)25.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Same tariff as DeepSeek-V4-Flash; input prices are cache-miss. Peak hours are 01:00–04:00 and 06:00–10:00 UTC and off-peak is half the peak rate. Images are billed as tokens, at most 384 tokens per image; Files API uploads are free.

Pricing source ↗

// Strengths

• Adds image understanding to the V4 line without giving up V4-Flash's text quality on agent, reasoning and knowledge tasks
• Priced identically to V4-Flash, with images capped at 384 tokens each regardless of their original resolution
• Keeps the full 1M-token context and 384K-token output ceiling, so long documents with embedded figures fit in one request
• Three image-input paths (base64, external URL, Files API) plus a free Files API for reusing the same image across requests
• Reachable from the Chat Completions, Messages and Responses APIs, with DeepSeek Harness 0.1.1 support out of the box

// Best for

• Multimodal agent loops that must read screenshots, dashboards or UI state to decide the next action
• Chart, diagram and document understanding at long context (DeepSeek reports 64.3 on Chartography)
• Front-end and design-to-code work where the prompt is a mockup image plus a repository
• Batch image analysis where the same reference images are reused across many requests through the Files API

// How to access

// DeepSeek V4 — every version

The full lineage of the DeepSeek V4 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.