跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

DeepSeek's First Vision Model Gets a Real Multimodal Stress Test

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

DeepSeek's first vision model is fast, cheap, and already available through its harness tooling, but the over-inference pattern means any developer piping its output into a downstream agent or user-facing feature needs a verification layer. The model will confidently tell you where the elevator is, even when the photo shows no elevator at all.

Summary

DeepSeek's experimental vision model, V4-Flash-Vision-Exp, was put through four multimodal reasoning challenges that went far beyond simple image captioning: a detective's desk puzzle, a dense supermarket shelf, a train station navigation table, and a parking lot spatial-relations task. The model handled OCR, object counting, and structured data extraction reliably, and it even spotted contradictions across multiple clues in the detective scenario. But a pattern emerged across every test: when the image lacked explicit real-world detail, the model confidently fabricated it, turning a supermarket receipt into proof of a person's location or inventing elevator directions from a map that showed none.

The parking-lot challenge, which required strict rule-based spatial reasoning, ran for 48 steps and over ten minutes in DeepSeek Harness's minimalist mode, yet produced the most stable result of the four. The station navigation task, by contrast, triggered the most aggressive hallucination, with the model generating turn-by-turn walking directions and accessibility advice from a static information board. The total API cost for the entire battery of tests came to ¥2.50, making the experiment cheap enough to run repeatedly while the model is still in its experimental phase.

Takeaways
DeepSeek-V4-Flash-Vision-Exp is an experimental release, accessible now through the DeepSeek Harness web UI and API.
Basic image analysis completes near-instantly, with multi-angle descriptions generated in a few steps.
The model failed to identify a partially visible charger plug in a photo, a detail that Gemini caught.
Across four structured reasoning challenges, OCR, object counting, and math calculations were largely accurate with minimal hallucination.
In a detective-style desk puzzle, the model correctly linked multiple clues and spotted a timeline contradiction, but it upgraded a receipt into proof of a person's physical location, which is over-inference.
In a train station navigation task, the model fabricated elevator locations, stair warnings, and directional distances that the source image never provided.
A parking-lot spatial-reasoning test ran for 48 steps and over ten minutes, yet produced the most stable and rule-compliant output of all four challenges.
The entire test session cost ¥2.50 in API fees.
Copy-pasting images into the DSH web UI via Command+C/V was broken; drag-and-drop and the @ method worked.
Conclusions

Most published 'tests' of new vision models stop at single-image captioning, which exercises perhaps a tenth of a model's actual multimodal reasoning surface. Structured, multi-question challenges reveal failure modes that simple recognition never touches.

The model's over-inference is not random noise; it follows a consistent pattern of completing partial real-world cues into full narratives. A receipt implies a person was present; a map implies navigation infrastructure. This is the same class of hallucination that makes LLMs dangerous in high-stakes domains, now extended into the visual modality.

Spatial reasoning with explicit, rule-bound constraints (parking spaces, EV designations, exit paths) produced better results than open-ended real-world inference (train stations, detective clues). The model performs better when the answer space is closed and the rules are visible in the image itself.

Running a vision model through an agentic harness that takes 48 internal steps for a single parking-lot query suggests the model is doing heavy internal chain-of-thought work. The cost stayed at ¥2.50, which implies the per-step pricing is low enough to make this viable for batch processing, but the latency makes it unsuitable for real-time use.

Concepts & terms
Multimodal reasoning
The ability of a model to process and reason across multiple types of input — here, images combined with text prompts — going beyond simple object recognition to tasks like spatial planning, timeline construction, and contradiction detection.
Over-inference (hallucination in vision models)
When a vision model generates details not present in the image, such as inferring a person's location from a receipt or inventing elevator directions from a map that shows none. It is the visual equivalent of an LLM fabricating facts.
DeepSeek Harness (DSH)
DeepSeek's agentic runtime that wraps model inference in a multi-step reasoning loop. In minimalist mode, it can run dozens of internal steps to solve a single query, as seen in the 48-step parking-lot test.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗