Google Gemini 3 Update: A Multimodal Ambush Before DeepSeek V4

The internet is waiting for DeepSeek V4 (or the rumored R2 reasoning model), but Google clearly doesn't want this window to be too quiet. Early this morning Beijing time, Google DeepMind updated its Gemini 3 Deep Think mode.

This update didn't continue the traditional text benchmark race but instead showcased a capability more akin to an "engineer": Physical World Reasoning.

From Sketch to Object

If DeepSeek V4 is expected to maximize code and logic, the new Gemini 3 Deep Think has chosen a different path: Physical Intuition.

Google demonstrated a core feature—"Multimodal Engineering Reasoning." A user draws a crude sketch on a napkin, and Gemini 3 Deep Think not only identifies the intent but also directly generates an STL file ready for 3D printing, even correcting structural mechanics hazards along the way.

This is indeed interesting. Previously, we compared who was better at solving problems (math/code); now Google has changed the subject to "understanding physical space."

Pressure on DeepSeek

The market has high expectations for DeepSeek's next-generation model. Rumor has it that DeepSeek V4 will possess a 1 million+ context window, potentially further improving code generation capabilities, and even optimized to run on consumer-grade graphics cards (like the RTX 4090). The R2 version is believed to attempt new boundaries in pure logical reasoning using reinforcement learning.

The update to Gemini 3 Deep Think adds a threshold to the upcoming competition: for current flagship models, just writing code might not be enough; they also need to understand diagrams and the physical constraints of the real world.

Google's Technical Moat

This update mainly reflects two directions:

1. Combination of Vision and Logic: It's no longer simple "image captioning" but "image modeling." This requires the model to have a deep understanding of spatial relationships.
2. Rigor in Scientific Applications: In International Mathematical Olympiad (IMO) level tests, the new Gemini 3's performance approaches that of human gold medalists, beginning to attempt to solve actual scientific problems.

The Competition Has Just Begun

For developers, competition at this level is always good news. Google has played its card, showing the upper limit with "physical reasoning"; next, we watch how DeepSeek responds. Will it continue to maximize efficiency and the open-source ecosystem, or will it bring out a new killer app in reasoning capabilities?

The first AI showdown after the 2026 Spring Festival is set between Google and DeepSeek.

---

👉 Follow the latest updates and DeepSeek V4 tests at our official site: deepseekv4.app