刚刚，DeepSeek 大升级，V4真的不远了

Just Now: DeepSeek Major Upgrade — V4 Is Truly Around the Corner

AppSo in focus

From WeChat public account: APPSO, by the team behind “Discover Tomorrow’s Products”. Original title: “DeepSeek Major Upgrade, V4 is Not Far Away | Experience Details”

Moments ago, DeepSeek’s web interface received a significant update.

No press conference, no blog post, not even a single official tweet. Two new icons have appeared above the input box on DeepSeek’s web — a lightning bolt and a diamond, representing “Fast Mode” and “Expert Mode” respectively.

Hovering reveals the prompts: Fast Mode “suitable for daily conversations, instant response”, Expert Mode “good at complex problems, wait times during peak”.

Based on tests and community analysis, differences are roughly as follows:

Fast Mode can recognize text in images and files, responds quickly. The trade-off is that it likely runs a lighter V4 Lite model optimized for speed.

Expert Mode appears to route to a larger, stronger model — likely some form of the upcoming DeepSeek V4. But currently it does not support file uploads nor multimodal capabilities. Wait, a stronger model with fewer features?

We also ran simple tests.

For instance, asking both modes to write a p5.js program simulating a ball bouncing inside a rotating hexagon under gravity and friction. Expert Mode produced more physically intuitive results — better trajectory accuracy and realistic bouncing.

In contrast, Fast Mode’s result was visibly inferior.

This gap illustrates the point: physics simulation demands strong mathematical reasoning. A weaker model tends to produce “physics-like but wrong” results. Expert Mode’s performance here reflects tangible ability differences.

But netizen @AiBattle_ got surprising results when building a Space Invaders game: output gap between Expert and Fast Modes was not obvious.

The tester speculated: “I suspect Expert Mode is still routing to some version of V4 Lite. The full V4 might take a while to land on the web.” That aligns with external reports — LatePost reported V4 official version expected around April this year, likely still open-source SOTA, but “not overwhelmingly stronger”.

In other words, this grayscale “Expert Mode” may not be the final form.

Creative writing: we gave both a debate topic — “In defense of boredom: argue that boredom is a luxury for modern people”.

Expert Mode produced longer output with more complete logic chains; Fast Mode felt more natural and plain.

Interestingly, speed difference was minimal on this task; Expert Mode even had shorter thinking time. Possibly because creative writing is less sensitive to model scale than mathematical reasoning.

For simple tasks, both modes perform similarly; the deeper the reasoning needed, the more Expert Mode shines.

For the math puzzle “If you add 1 meter to a rope around Earth’s circumference and lift it evenly, how high is the gap?” Both gave the same answer but with different processes. Fast Mode was brief; Expert Mode broke down each step, clearly documenting reasoning.

Notably, only Fast and Expert modes are live now, but earlier leaks suggest a third option — “Vision Mode” — is on the way.

Teortaxes, a blogger focused on DeepSeek’s tech roadmap, believes making Vision a separate category is unusual. DeepSeek previously refused to deploy DS-VL series on the web citing “immaturity”. If Vision Mode goes live, it might be backed by a fully functional VLM.

Teortaxes goes further — this vision model might not be a conventional VLM but some “deep unified world model”, an evolution of the Janus series or another non-traditional architecture.

Of course, that’s just speculation. Another user @xhyctf says reverse-engineered frontend code shows Vision Mode has no independent model; “visual understanding” is just a parameter — filefeature.vision=true — added under Fast Mode.

Still, one thing is certain: DeepSeek has long been making multimodal moves, maybe just waiting for the right timing. Putting Fast/Expert modes in front of users signals a bigger direction:

DeepSeek begins product tiering.

Since its breakout early last year, DeepSeek’s product logic has been highly “anti-commercial” — low API pricing, completely free web, no functional barriers. But problems follow: long-term “all-free, no-tier” operation is commercially unsustainable.

Now, DeepSeek is creating two entry points for “stronger model” vs “basic model”. Expert Mode remains free for now, but once the architecture is set, building a paid system on top becomes technically straightforward.

Of course, tiering isn’t necessarily just about charging fees.

Naturally routing users — daily conversations to Fast Mode, deep reasoning to Expert Mode — is itself a compute scheduling strategy, limiting quotas and easing peak pressure.

Charging is one path, quotas are another — both viable, possibly in parallel. Looking at the big picture, this is a well-prepared roadmap:

Grayscale tiered entry → let users perceive differences → integrate multimodal & file capabilities → open Vision Mode → set quotas / price for more powerful modes. This is personal speculation; DeepSeek never plays by the book.

Admittedly, people have tagged DeepSeek with many labels — tech idealism, anti-commerce, inclusive AI. Yet GPU inference costs are real every month. However lucrative High-Flyer’s quantitative returns, it’s hard to indefinitely cover a global AI service’s free operation through API revenue alone.

DeepSeek has stirred up the entire AI circle, but reality will eventually stir DeepSeek as well.