Gemini 4 Argon writes 27x PyTorch-speedup GPU megakernel, ranks 3rd on KernelBench-Mega
scaling01 · x · 2026-10-14
On KernelBench-Mega, Gemini 4 Argon [high] wrote a one-launch Kimi-Linear Decode megakernel for the RTX PRO 6000 hitting 27.0x the optimized PyTorch baseline — 3rd of 35, behind Opus 5.5 (35.5x) and Sonnet 5.5 (32.1x); Gemini 3.8 Flash managed only 2.7x. It climbed the same optimization curve as Sonnet 5.5 for three hours before the session was cut off when Gemini's safety filter blocked one of its own profiling requests.
More from Models
- Expand Fetch launches: agents see the web, hits 74% vs Exa's 55% on hard extraction benchmark — bclavie · 2026-10-15
- Blogger Starts Testing Rumored Unreleased Gemini 4 Argon Model — _arohan_ · 2026-10-14
- Security researcher banteg finds Argon model reasoning about his fame during audits — banteg · 2026-10-14
- OpenAI models write good code but take longer paths; Sol 6.1 is the best worker — haider1 · 2026-10-14
- New Head-to-Head: Which AI Model Builds the Best Lego iMac From Accurate Parts? — dreamwieber · 2026-10-14
- JetBrains Mellum 2.1 12B MoE impresses in local tests: 140 tok/s on a 16GB GPU — Shadow_s_Bane · 2026-10-14