HIGER|11月 20, 2026 14:56
Google has made rapid progress in large-scale models recently, not just with another new version release, but also with new advancements in model capabilities, search access, and commercial usage scale.
1. Gemini 4 has built a set of models tiered by task and cost. Carbon is designed for more complex inference and long-process tasks, Hydrogen is for faster and cheaper everyday use, and Neon, released on November 18th, is for complex engineering, analysis, and research.
2. Hydrogen is now available in Google Search globally, as well as in both free and paid Gemini applications. Google's advantage lies not only in the models themselves, but also in the readily available distribution channel of search, allowing model upgrades to directly reach ordinary users.
3. Neon's independent benchmarks are approaching the forefront. In the Vals Index, which covers practical work tasks in finance, programming, and law, Neon ranks second, close to first place, while the benchmark records lower testing costs. However, this does not necessarily mean that Google has achieved a comprehensive lead.
4. Usage scale is also expanding. Google's Q3 earnings report stated that Gemini app users reached 1 billion, and the first-party model API processed approximately 32 billion tokens per minute, a 40% increase from the previous quarter. Google Cloud revenue grew by 103% year-over-year during the same period, but this includes infrastructure, other cloud services, and TPU system sales, which cannot all be attributed to the large model.
5. Multimodal computing is expanding from generating content to generating interactive environments. Genie 4 is now available to eligible Ultra users and paid APIs, allowing the generation of interactive scenes using text or reference images, along with simultaneous sound generation. While it still faces challenges with complex physical interactions, latency, and long-term scene stability, it's no longer just about generating a video for people to watch.
Previously, SemiAnalysis published a report pessimistic about Google's large model, but due to recent developments, SemiAnalysis has had to apologize to Google.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink