DeepSeek Image Recognition Mode Goes Live: Completing the Final Piece of the Multimodal Puzzle

DeepSeek multimodal researcher Xiaokang Chen confirmed today (June 18) that the Image Recognition Mode has officially launched on both web and app platforms. This feature, dubbed “Whale Opening Its Eyes,” has progressed from a beta rollout at the end of April to full public release.

It’s worth noting that this launch makes DeepSeek the last major domestic model vendor to complete its visual understanding capabilities. Alibaba Qwen-VL, Zhipu GLM-V, ByteDance Doubao, Moonshot Kimi, and MiniMax Hunyuan had all completed their multimodal setups earlier. DeepSeek arrived later, but not too late — judging by its product cadence, it was waiting for a sufficiently mature technical solution.

Not Just OCR, but True Visual Understanding

Image Recognition Mode now sits alongside Fast Mode and Expert Mode as the third primary entry point in DeepSeek’s main product. The placement itself sends a signal: visual understanding is not a supplementary feature, but a standalone capability line.

Testing shows that the capabilities far exceed simple text extraction. For example, a user uploaded a photo of a person and asked, “What pose is this?” After 8 seconds of reasoning, DeepSeek’s answer included positional analysis, arm posture, head orientation, hair fall status, clothing style, and light-shadow contrast. It ultimately concluded this was a “laid-back reclining pose” or a “cool atmosphere pose,” often seen under certain tags on Xiaohongshu or Douyin.

This combination of structured breakdown + self-review + cultural context recognition has moved beyond traditional OCR capabilities. During its reasoning process, DeepSeek also lists other possible interpretations (“lady-like pose,” “melancholy selfie”), then self-corrects to determine the final answer. This approach is closer to how humans understand images: observe details first, then synthesize a verdict, and finally integrate cultural background for interpretation.

In contrast, earlier visual language models were more like “describe the picture” — identifying objects, colors, and spatial relations. DeepSeek's Image Recognition Mode goes further, attempting to understand the intent, atmosphere, and cultural symbols behind an image. This offers much greater value in practical use cases: designers can analyze competitor visual styles; operators can extract emotional tags from images; developers can analyze UI screenshots.

Four Months of Preparation, Clear Product Cadence

The launch of Image Recognition Mode didn’t come out of nowhere. Looking back, DeepSeek began preparing for this feature in early April.

On April 8, icons for Fast Mode and Expert Mode first appeared above the web input box. This was the first time DeepSeek layered capabilities in the product interface. Some tech KOLs noted that separating Vision as its own category was unusual, implying that if it launched, it would likely be backed by a fully functional VLM (visual language model).

On April 24, the preview of DeepSeek-V4 was released, featuring million-token long context, Agent capability, and reasoning performance. However, V4 itself lacked native multimodal support, which was considered a notable shortcoming.

On April 28, DeepSeek’s multimodal lead Xiaokang Chen posted an image on X platform: two blue whales — the left wearing a black blindfold with “XX” drawn on it, the right removing the blindfold to reveal its eye. This was widely interpreted as “Whale Opening Its Eyes,” hinting at the imminent arrival of visual capabilities.

On April 29, Image Recognition Mode entered beta testing. Some users saw the third label in web and app with a tooltip saying “Image understanding feature in beta.” That same day, a developer on V2EX noted that the DeepSeek API was returning an “Image Recognition Mode” field, though it was not yet accessible externally.

From April 8’s layered interface, to April 24’s V4 text-only release, to April 28’s teaser, and finally April 29’s beta — this was a continuous product rhythm. DeepSeek didn’t rush to release a half-finished product; it waited until the technical solution was sufficiently mature.

Now, a month and a half after beta began, Image Recognition Mode has been fully rolled out as of June 18. The timing is noteworthy — completing multimodal capability before mid-year sets the stage for Agent scenarios and more complex applications in the latter half of the year.

Technical Foundation: From DeepSeek-VL to Janus-Pro

DeepSeek is no newcomer to multimodal research. As early as 2024, it launched the DeepSeek-VL series models aimed at real-world visual language understanding, covering charts, webpages, formulas, scientific literature, and natural images.

Later it rolled out the Janus series, attempting to unify multimodal understanding and visual generation. Janus-Pro outperformed DALL-E 3 and Stable Diffusion 3 on the GenEval image generation benchmark and was one of the most noteworthy achievements in the open-source multimodal space at the time.

In April this year, DeepSeek revealed a core framework called “Thinking with Visual Primitives.” The core idea: instead of having the model directly “look” at the entire image, first break it down into a series of visual primitives (basic visual elements), then let the model reason based on those.

This design has two benefits:

1. Reduced computational cost: processing visual primitives is far less costly than full images, particularly advantageous in multi-step reasoning scenarios.
2. Improved interpretability: the model’s reasoning can be traced back to specific visual elements rather than opaque outputs.

Based on testing, this framework has already been implemented in the product. DeepSeek explicitly lists elements like “torso angle,” “arm position,” and “hair state” during reasoning, then synthesizes them for judgment. This structured reasoning is exactly what “Thinking with Visual Primitives” entails.

It’s worth noting that there have been personnel changes in DeepSeek’s multimodal team recently. Key contributor Chuong Ruan joined autonomous driving company Yuanrong Qixing as chief scientist in April, and Wei Haoran, core author of DeepSeek-OCR series, left around the Lunar New Year. However, judging by product progress, these changes haven’t affected the pace of multimodal capability advancement.

Filling the Gap or Getting the Ticket In?

With Image Recognition Mode now live, all major domestic model vendors have completed their multimodal setups:

• Alibaba Qwen: multiple generations of Qwen-VL visual language models covering general understanding, OCR, chart analysis.
• Zhipu GLM: GLM-V spanning GLM-4 and GLM-5, supporting image and video understanding.
• ByteDance Doubao: all-modality core positioning covering text, images, audio, and video.
• Step: Step 3.5 Flash open-sourced all-modality capabilities.
• Moonshot Kimi: parallel development of visual understanding and coding.
• MiniMax: differentiated advantage in video generation models.
• Tencent Hunyuan: Hy3 preview compared to DeepSeek and Alibaba.

In this list, DeepSeek is the last to complete visual understanding — but “last” doesn’t mean behind; it may be a strategic choice.

DeepSeek’s core advantages lie in reasoning capability and long context. The R1 series’ reasoning performance is close to GPT-4, and V4’s million-token long context costs are among the lowest in the industry. These two abilities are its moat and market differentiators.

Before firmly establishing these two core strengths, DeepSeek had no need to rush into multimodal. While multimodal is important, launching prematurely could dilute reasoning capability and push long-context costs out of control — the proverbial “picking up pennies and losing dollars.”

Now with Image Recognition Mode live, DeepSeek has sufficient technical reserves in reasoning, long context, and multimodal. Completing multimodal now is more like securing an entry ticket for the next phase of competition rather than simply plugging a gap.

The reason is simple: Agent scenarios require visual understanding.

In the Agent Era, Visual Understanding is Infrastructure

In V4’s release notes, DeepSeek explicitly mentioned optimizations for mainstream Agent products like Claude Code, OpenClaw, OpenCode, and CodeBuddy, making Agent capabilities one of V4’s three core selling points.

But purely text-based Agents have limited ceilings. When Agents need to operate browsers, read screenshots, view dashboards, recognize UI elements, and process charts and PDFs, lacking visual understanding prevents closed-loop operation.

Zhipu’s AutoClaw, Alibaba Cloud’s Coding Plan, Anthropic’s Computer Use — all share a logic: for Agents to truly enter productivity scenarios, visual capability is infrastructure, not a luxury.

Some concrete scenarios:

• Automated testing: Agent must read UI screenshots, determine button positions, text content, and layout correctness.
• Data analysis: Agent must understand chart types, axis meanings, and data trends — not just handle raw numbers.
• Document processing: Agent must parse PDFs, scans, and handwritten notes to extract structured information.
• RPA scenarios: Agent must recognize desktop app UI elements and simulate human operations.

In these scenarios, visual understanding is not optional — it’s essential. If DeepSeek wants a foothold in the Agent market, image recognition capability must be present.

From this perspective, the launch of Image Recognition Mode isn’t just about filling a gap — it’s about paving the way for Agent use cases. With reasoning and long-context advantages already in place, adding visual understanding completes the triad for a full Agent solution.

Next Steps: Video Understanding and Multimodal Generation?

Currently, Image Recognition Mode supports only image understanding and hasn’t opened video understanding or image generation capabilities. But given DeepSeek’s prior technical work, both directions are likely in preparation.

Janus-Pro has already proven its superiority over DALL-E 3 for image generation tasks, validating the technical approach. Video understanding was also present in the DeepSeek-VL series, though not yet in the main product.

Looking at other vendors’ trajectories, video understanding and multimodal generation will likely be DeepSeek’s focus in the second half of the year. Video understanding in particular sees high demand in short video, live streaming, and surveillance analysis — a much larger market than still images.

Another point of interest is API access. The DeepSeek API documentation doesn’t yet list Vision, Image Input, or related calls, meaning image recognition isn’t available via API. For developers, API access directly impacts integration cost and applicability.

If DeepSeek can complete API access for Image Recognition within the year, alongside video understanding and multimodal generation, its competitiveness in the multimodal space will rise significantly.

In Conclusion

The launch of DeepSeek Image Recognition Mode marks the completion of multimodal capability setups among major domestic large model vendors. In four months, DeepSeek has moved from pure text reasoning to full text-image capability.

This “gap-filling” wasn’t mere feature stacking — it waited for a mature technical solution. From early April’s product layering, to late April’s beta testing, to mid-June’s full release, DeepSeek’s product cadence has been clear and deliberate.

The launch of Image Recognition Mode isn’t an endpoint, but a starting line. Visual understanding + reasoning capability + long context combine into a complete Agent solution. DeepSeek’s next step is to run this capability suite in real-world scenarios, proving it’s not just a tech demo but a true productivity tool.

From a technical reserve standpoint, DeepSeek holds the cards. From a product pacing standpoint, it advances methodically. The question now isn’t whether it can do it — it’s how quickly and how well it will be done.

---

References

• DeepSeek Image Recognition Mode Officially Launched on App and Web — IT Home — IT Home’s initial report on the launch of Image Recognition Mode
• Other sources have been integrated but not listed due to domain restrictions