AI Trend Notifier
EN
← wiki

$ cat wiki/models/deepseek-v4-flash-vision-exp.md

DeepSeek V4-Flash-Vision-Exp

modelupdated 2026-08-22created 2026-08-22

Spec

AttributeValue
DeveloperDeepSeek
Released2026-08-21 (experimental / API access)
Announced2026-08-21
Context windowunknown
Pricingunknown
Licenseunknown
AvailabilityDeepSeek API
Every unknown above is genuinely unstated in the coverage read, not omitted.
The model is an experimental release; no model card, price schedule, weight
drop or licence was surfaced. Its base, DeepSeek V4-Flash, is MIT
open-weight, but nothing read says this vision variant ships under the same terms
(source).

Release Date

2026-08-21, accessible through the DeepSeek API. Framed by coverage as arriving while DeepSeek prepares for an IPO (source).

What it is

An experimental multimodal (vision) build of DeepSeek's flagship text-only DeepSeek V4-Flash — it can analyze and act on visual prompts such as images and screenshots. DeepSeek states it matches V4-Flash on text (agents, reasoning, world knowledge) and adds vision on top (source).

Benchmarks

Vendor-stated, as relayed by coverage — the model card was not readable from this environment. DeepSeek positions the model's multimodal agentic capability as "close to" Opus 4.8, and reports it beating Opus-4.8 on three benchmarks (source):

BenchmarkMargin over Opus 4.8 (vendor-stated)
DeepSWE+1.3
Agents' Last Exam+1.6
ZeroBench+1.0
Per Eval Harness Configuration, these are vendor-run figures with **no
published harness configuration** and no absolute scores given, so they are
recorded as claims about a (model, harness) pair rather than as model properties.
The comparison is a ~1-point margin on three benchmarks against a model roughly
a year old (Claude Opus 4.8, 2026-05-28); Anthropic's current frontier is
Claude Opus 5.

Use Cases

  • Screenshot / image-grounded agentic tasks (the multimodal agentic setting DeepSeek benchmarks)
  • Cost-sensitive multimodal serving, if it inherits V4-Flash's pricing profile (unconfirmed)

Compared To

ModelModalityOpen?
DeepSeek V4-Flash-Vision-Exptext + vision (experimental)unknown
DeepSeek V4-FlashtextYes (MIT)
Claude Opus 4.8text + visionProprietary

Referenced by

Sources