DeepSeek Finally Got Eyes

DeepSeek V4-Flash was already the most cost-effective coding model on the market. It had one gap: it couldn't see. Users kept asking it to analyze screenshots, and it kept hallucinating vision capabilities — inventing tools to describe images it couldn't actually process.

Today that gap closes. DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to the V4-Flash architecture. Combined with the new Files API, this is the first time DeepSeek's agent-optimized line can actually see what it's working with.

What Ships

The release is three things bundled together:

1. The vision model itself. Same V4-Flash text backbone, but accepts images as input. JPEG, PNG, GIF, WebP. Images get tokenized at up to 384 tokens each, then billed at standard V4-Flash rates. The announcement claims it "matches V4-Flash on text capabilities" — agents, reasoning, world knowledge — and brings multimodal agent performance "close to Opus 4.8."

2. An 800×800 resolution cap. Images are automatically scaled down so the total pixel count is roughly that of an 800×800 image. This is the single most important technical detail in the entire release. More on why below.

3. The Files API. A free companion service: upload an image once, reference it by file_id across requests. Saves bandwidth on repeated analysis workflows. Small quality-of-life win, but it matters for agent loops that need to reference the same screenshot multiple times.

The Numbers That Matter

The HN thread surfaced the key benchmark: DeepSWE 59.3%. That's within spitting distance of 5.6-Sol Medium (61% ± 2%) — but at roughly 1/18th the cost. The cost-performance ratio is the headline here.

V4-Flash was already the budget king for agentic coding. Now it sees. The cost delta against Sol Medium, Opus, or Sonnet is so large that even if vision quality is merely "good enough" for 80% of use cases, the economic case is overwhelming.

graph TD
  A[V4-Flash-Vision-Exp] --> B{Multimodal Agent}
  B --> C[DeepSWE 59.3%]
  B --> D[Cost: ~1/18th of Sol Medium]
  C --> E[vs Sol Medium: 61% ± 2%]
  C --> F[vs Opus 4.8: close per DeepSeek]
  D --> G[$0.15/M input tokens]
  D --> H[384 tokens per image]
  style C fill:#1a3a5c,stroke:#3b82f6
  style D fill:#1a3a1a,stroke:#22c55e

Where It Falls Down

HN users found the rough edges fast.

The 800×800 limit is genuinely limiting. For OCR, reading full A4/letter-sized documents, or analyzing high-resolution screenshots, that's a hard ceiling. You can split images and feed them in tiles, but that adds complexity and cost. As one commenter put it: "Oof 800 by 800 kills a lot of use cases."

Fine-grained visual reasoning is weak. The model fails the "clock test" — a simple analog clock image that Qwen 3.8 27B gets nearly right. It also misidentifies landmarks, confusing Wells Cathedral for Salisbury Cathedral on a test ByteDance Seed 2.1 Turbo handles correctly.

Tool call results can't be images. As one commenter noted on HN, for agentic use cases the main value of vision is "giving the model a tool to take screenshots and verify its work." This release doesn't support images as tool call result types — it's oriented toward human-prompted image analysis, not autonomous agent vision loops.

Context: The Hallucination Era

The funniest context for this release is that V4-Flash (0731) was notorious for pretending it had vision. Multiple HN commenters reported the same pattern: they'd ask it to analyze a screenshot, and it would go through the motions — pulling screenshots off devices, describing pixel data — before realizing it actually couldn't see anything.

This is a real problem in the field. Models trained on multimodal data learn that "vision" is a thing models do, and when they're text-only, they confabulate the capability. DeepSeek's fix is to actually give them the capability. Simple, but effective.

The Bottom Line

V4-Flash-Vision-Exp is the cheapest multimodal agent model on the market, by a wide margin. If you're building agent workflows that occasionally need to look at a UI screenshot, read a simple chart, or verify a visual output, this is the most cost-effective option available today.

The 800×800 resolution cap and weak fine-grained visual reasoning mean it won't replace Sonnet or Opus for serious document analysis or detailed visual tasks. But for the most common agentic vision use cases — "did my deploy succeed?" screenshot checks, "what does this error page look like?" — it's good enough, and it costs 18x less than the competition.

The Files API is a nice bonus. The DeepSeek Harness 0.1.1 integration is a signal of where they're pushing. And the fact that the model was hallucinating vision before it could actually see is the kind of story that makes you realize how weird this whole AI moment really is.