DeepSeek-V4-Flash-Vision-Exp

DeepSeek·2026-08-21

LLMsmultimodalcloud

Overview

Experimental vision variant of DeepSeek-V4-Flash-0731, live on the DeepSeek API platform as model 'deepseek-v4-flash-vision-exp'. Same 284B total / 13B active MoE and 1M-token context as the text build, extended to image input (text output only); images are billed as up to 384 tokens each and can be passed as base64 through Chat Completions, Messages and the Responses API. DeepSeek describes text capability — agents, reasoning, world knowledge — as unchanged versus V4-Flash and reports the gains on agent benchmarks that require visual understanding: Terminal-Bench 2.1 83.9 and DeepSWE 59.3. Unlike the rest of the V4 line this is a closed API release — no open weights, no Hugging Face repository at launch, and the 'Exp' tag marks it as experimental.

Capabilities and innovations

284B total / 13B active MoEImage input, text output1M token context windowMultimodal agent tasksChat Completions, Messages and Responses APIImages billed at up to 384 tokens eachFirst vision-enabled model of the DeepSeek V4 lineVision added on top of the V4-Flash-0731 checkpoint at unchanged parameter countAPI-only release — first V4 build without open weights

Benchmarks

Benchmark	Score	Source
Terminal-Bench 2.x	83.9%	vendor Terminal-Bench 2.1 83.9 from the launch announcement (DeepSWE 59.3 in the same table, no schema field). Version 2.1, comparable to the 79.0 recorded for V4-Flash-0731 but not to the 2.0 figures of the April preview. Announcement not fetchable from this environment; value taken from consistent secondary transcriptions and not independently reproduced.

GPQA Diamond: not reported by the vendor — DeepSeek states text capability is on par with V4-Flash-0731 but publishes no separate run for this build. The 0731 values are deliberately not carried over — a claim of parity is not a measurement.
Humanity’s Last Exam (no tools): not reported by the vendor — Same parity claim as GPQA, no separate run published.
SWE-bench Pro: not reported by the vendor — The launch note reports agentic suites (Terminal-Bench 2.1, DeepSWE) only.
MMLU-Pro: not reported by the vendor — Same agent-focused launch note.
MMMU-Pro: independent evaluation pending — This is the first DeepSeek model where MMMU-Pro would apply, but no multimodal reasoning benchmark was published at launch and no independent run exists yet.

API pricing

$0.22 per million input tokens, $0.66 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-09-01)

Architecture and hardware

Parameters: 284B total, 13B active per token (MoE)

More from DeepSeek

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.