Testing Gemini 4 Hydrogen and Argon vs. Gen 3 over 1,474 Multi-Turn Queries

news.ycombinator.com4 min read

In a technical report analyzing 1,474 multi-turn queries, researchers used an unmodified C++ native telemetry engine to evaluate next-generation LLMs including Gemini 4 Hydrogen and Gemini 4 Argon against Generation 3 models. The study uncovered contrasting failure geometries, decode-stage instabilities despite high implicit caching, and significant drops in thinking tokens. The findings highlight the limitations of prefill caching and the emergence of severe runtime behaviors under heavy compute constraints. Open article →

More on this topic