DeepSeek V4 Flash Vision Exp

DeepSeek's official news post dated September 9, 2026, introducing DeepSeek-V4.1-Flash, explicitly addresses the disposition of DeepSeek-V4-Flash-Vision-Exp by stating that V4-Flash and V4-Flash-Vision-Exp are retired and that, for compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V Beyond the retirement notice, the page focuses on V4.1-Flash capabilities (552B-parameter MoE with 8B/16B active parameters, new Causal Encoder-Decoder architecture, reduced KV cache footprint, and API pricing changes effective September 10, 2026), and outlines that V4-Pro requests will also route to V4.1-Flash startin

DeepSeek announced the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp on August 21, 2026 via its first-party API documentation news page. The release matches V4-Flash on text capabilities — including agents, reasoning, and world knowledge — while making a major leap on multimodal agent benchmarks, with repo The release also opens a Files API for free image uploads reusable across requests. Images are billed at up to 384 tokens each at V4-Flash pricing, and the API supports Chat Completions, Messages, and Responses endpoints. Image input accepts base64, external URLs, or Files API references, with the official vision guide

MindStudio's analysis provides the most substantive technical detail on DeepSeek V4 Flash Vision Exp, describing the architecture as visual modules added to the existing V4 Flash base with continued training, rather than a from-scratch multimodal build. The model has reportedly been released as open weights and accumul The analysis surfaces two important caveats for developers. First, a methodological footnote in DeepSeek's own numbers reveals that the 0731 baseline on ApexBench and Agents' Last Exam ignored multimodal input entirely, so those deltas are not apples-to-apples comparisons. Second, Opus 4.8 still leads on most text-agen

An NVIDIA Developer Forums community thread titled "DeepSeek v4 Flash Vision Exp is Released as Open Weights" (dated September 1, 2026) documents local-inference deployment efforts for DeepSeek-V4-Flash-Vision-Exp on DGX Spark / GB10 hardware. Forum users confirm native vision support while keeping text and agent perfo Thread participants identify the main blocker for single-DGX-Spark users as the current approximately 168 GB FP8/FP4 checkpoint and the lack of native Vision-Exp support in ds4/vLLM, with the community awaiting a suitable single-Spark quantization (roughly 80–100 GB) and a vision encoder/aligner port. Shared tool-eval-

DeepSeek announced V4 Flash Vision Exp on August 21, 2026, as an experimental multimodal extension of the V4 Flash line, adding native image understanding while preserving text and agent capabilities. According to DeepSeek's official announcement cited in the analysis, multimodal-agent performance moves close to Claude Engineering caveats matter: the "Exp" suffix signals that identifiers, limits, and pricing may change, and benchmark methodology, production reliability, and weights status remain unverified. Builders are advised to run workload-specific tests before treating the model as a production default, particularly because repo

Edenai's comparison frames DeepSeek V4 Flash Vision Exp as winning 3 of 11 self-reported benchmarks against Claude Opus 4.8 and trailing by 1.1 to 12.0 points on the others. The strongest relative wins are on Agents' Last Exam (+1.6) and ZeroBench (+1.0), while Chartography (-0.7) and Terminal Bench 2.1 (-1.1) remain c The analysis also documents the evaluation configuration DeepSeek used for its own benchmarks, including Harness Minimal Mode with max tokens at the model ceiling, top_p 0.95, and temperature 1.0. No independent verification of these numbers exists, which is a meaningful caveat when interpreting the comparison since a

An independent benchmark aggregator page dated August 21, 2026, ranks DeepSeek-V4-Flash-Vision-Exp at position 27 on its composite LLM Stats Score, with a score of 46.0 and a blended price of $0.48 per million tokens. It lists per-benchmark ranks and scores including Terminal-Bench 2.1 at 0.84/1 (rank 15), DSBench-Hard The aggregator situates DeepSeek-V4-Flash-Vision-Exp between DeepSeek-V4.1-Flash (score 51.8 at $0.24 blended) and DeepSeek-V4-Flash-0731 (44.7 at $0.066) on its cost-efficiency chart, illustrating its premium pricing relative to the non-vision V4 Flash sibling. Scores and ranks are presented as sourced directly from t

DeepSeek's official API changelog confirms that DeepSeek-V4-Flash-Vision-Exp was released on August 21, 2026, as an experimental multimodal model accessed via model='deepseek-v4-flash-vision-exp'. The first-party change log lists its benchmark scores across Terminal Bench 2.1 (83.9), NL2Repo (57.7), DeepSWE (59.3), DSB The same first-party changelog entry, dated September 10, 2026, states that with the release of DeepSeek-V4.1-Flash, the previous-generation models V4 Flash and V4 Flash Vision Exp have been retired, and the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash. API pricing

DeepSeek's official API Change Log documents the August 21, 2026 release of DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal vision understanding model accessible by setting model='deepseek-v4-flash-vision-exp' on the DeepSeek API. The release entry lists benchmark scores including Terminal Bench 2.1 at 83.9, N A September 10, 2026 changelog update retires the previous-generation DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp, noting that the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to the newly released DeepSeek-V4.1-Flash, which is positioned as the smallest model in a new ar