I can't believe DeepSeek has just been up and running!
Recently, DeepSeek officially announced that DeepSeek-V4-Flash-Vision-Exp is an upcoming DeepSeek API platform.
In other words, DeepSeek's multi-modal model service is here.
According to the official tweet: Today, DeepSeek launched the multi-modal visual understanding model DeepSeek-V4-Flash-Vision-Exp on its API platform, and the developers set the model parameters in the request as deepseek-v4-flash-vision-exp.
DeepSeek Harness, which was just updated two hours ago, has also added a new model.
Interestingly, it was at this time that the hidden model Ox Alpha on OpenRouter and OpenCode, jokingly known as "Bullshit", sparked a heated debate on social media.
The self-introduction of this anonymous model with DeepSeek's release today is on almost the same set of keywords: code-oriented, long-range Agent working with real production environments, 1 million tokens in context, native support for text, image and video input, reasoning for compulsory activation and division of low, high, max into three tiers.
Ox Alpha is currently free, online in less than a day, and the volume of calls is so high that many people directly connect it to their smartphones.
It's supposed to be a domestic model, as the figure below suggests: the "visual" contrast between the two domestic models.
DeepSeek released a preview version of the V4 series on April 24 and has been open source for nearly four months.
Over the past four months, V4's other puzzles have been solved: The official V4-Flash API was unveiled on 31 July, and the official V4-Pro version and the official Agent framework DeepSeek Harness (dsh) have now landed on 178,000 GitHub Stars.
Visual is the last and the longest-awaited piece outside the community. So much so that the community can't wait to get their hands on the third-party visual encoder on V4-Flash-0731 for an unofficial "hybrid visual" version.
According to DeepSeek, V4-Flash-Vision-Exp is on par with the official version of V4-Flash in its pure-text capabilities, covering Agent, reasoning and world knowledge; while on the Agent Benchmark, which requires visual understanding, the multi-modal Agent capability is close to Claude Opus-4.8 compared to V4-Flash.
V4-Flash-Vision-Exp, V4-Flash-0731 and the eleventh benchmark comparison for Opus-4.8
Three examples of the capabilities given by the announcement are the creation of a Tibet autonomous travel custom PPT for high-net-worth clients under Agent, the second creation of DeepSeek Harness' Deep Sea Black Blueglass UI style, and the dynamic feature front end demo of a clay monster style.
Example 1: Generate a commercial customized Tibetan self-driving PPT under the Agent framework
Prompt: Help me do a Tibetan tactical ppt, Tibet South + Tibet North, a month, self-guided tour, I'm a high-end customized tour, only private service, different from other travel groups, core tuning: wild, original, exploratory, not routine playing routes, is a deeply human-level experience.
Scroll down and seeExample 2: Second creation of the official website of DeepSeek Harness
The model requires visual elements such as the deep blue sea, the glass UI, and the ASCII atomic pixel to be used.
Futurist-style developer website rebuilt after a long series of interactions
Example 3: Mini Demo on the front end of a pottery monster-style dynamic feature
The suggestion is to create models inspired by a group of cute 3D clay monsters joining a leaping retrospective dance pool party.
We've just talked about how to model, so let's try it out for ourselves.