In this video, I’ll be testing Google’s new Gemini 3 “Riftrunner” checkpoint on LM Arena, showing real prompts and results, how to try it yourself, and how it stacks up against X58, ECPT, and Sonnet.
—
Key Takeaways:
🚀 Google’s Gemini 3 Riftrunner checkpoint is live on LM Arena and looks like a release candidate.
🕹️ How to try: open LM Arena, select Battle mode, send prompts, and vote; Riftrunner will appear among models.
🧩 Visual tasks: floor plan is bland but passable; SVG panda with a burger is decent (burger great); Pokeball in ThreeJS looks awesome; butterfly garden is one of the best.
♟️ Chessboard with autoplay fails completely—the first Gemini 3 checkpoint to miss this test.
🎮 Minecraft Kandinsky clone runs but breaks on jump, sending the player roaming the sky.
💻 Coding: Rust CLI is good; Blender script is fine but lacks lighting/textures compared with X58.
📐 Reasoning: passes one math question and fails another; riddle solved (even generated an HTML page).
📊 Performance: roughly ~15% above Sonnet but ~14% below the best X58; overall ranks 5th among recent checkpoints.
🔒 Likely causes: extra safety filters, chat tuning, and/or quantization; could be a Flash-style variant, but unconfirmed.
🔮 Verdict: a solid step over current‑gen models, not a “3.5 Sonnet moment”; hoping for Ultra and access to X58 checkpoints.
📰 Rumors: Apple–Google talks suggest a 1.2T‑parameter model; speculatively Flash ~1.2T and Pro ~2T for faster speech/live use.
👍 Bottom line: good, not the best—please ship the final model instead of more checkpoints.
source
