Google Starts Gemini 4 Pretraining: Its Biggest Training Run Ever, Before 3.5 Pro Even Ships
Google quietly announced on July 21 that Gemini 4 has entered the pretraining stage — what the company calls its "most ambitious pretraining run yet." Meanwhile, Gemini 3.5 Pro still hasn't fully shipped.
Three Key Takeaways
Gemini 4 enters pretraining as a ground-up redesign. Google explicitly stated that Gemini 4 is not an incremental upgrade over 3.5, but a "completely revamped" foundation model. No details on architecture or parameter count have been disclosed, but the language suggests a full-scale training run from scratch rather than fine-tuning an existing model.
Gemini 3.6 Flash and specialized models launch simultaneously. 3.6 Flash targets lower inference costs and faster response times. Google also rolled out 3.5 Flash-Lite (a lightweight variant) and 3.5 Flash Cyber (optimized for cybersecurity workloads). These models are purpose-built for coding, reasoning, and security tasks, covering different price points and use cases.
Gemini 3.5 Pro remains stuck in partner testing. The model that was supposed to be Google's current flagship still hasn't reached general availability. Google says it's being tested with partners, suggesting it may not yet meet internal quality or safety benchmarks. One side is running late, the other side is getting a head start — Google's model cadence looks increasingly chaotic.
WangDou's Take
Google's model release cadence right now looks like a student juggling three assignments at once: the 3.5 Pro homework isn't turned in, the 3.6 Flash paper just got submitted, and 4.0 is already in rough draft. The generous read is that iteration is fast; the honest read is that 3.5 Pro's performance disappointed Google itself, so they're rushing to Plan B. From a competitive standpoint, Anthropic's Claude Opus 4 series and OpenAI's GPT-5.6 are already being used in production by enterprise customers — Google can't keep holding its spot with "coming soon." The real signal behind starting Gemini 4 pretraining may be that Google has concluded the current architecture has hit its ceiling and needs to be rebuilt from the ground up. If that's the case, 3.5 Pro is likely to become a transitional product that gets forgotten the moment Gemini 4 lands. That's the cruelest part of the AI arms race — your flagship hasn't even shipped, and the competition's next generation is already on the way.
Source: Droid-Life, AI Weekly, CNBC