Visit PhotoGenius AI: https://www.photogenius.ai/
In this video, I test the two new LM Arena checkpoints—Orionmist and Lithiumflow—rumored to be tied to Gemini 3, show you how to try them yourself, and compare their performance against ECPT and earlier checkpoints across 11 tasks (graphics, code, games, math, and reasoning).
—
Key Takeaways:
🚀 Google likely won’t launch Gemini 3 this week; two new LM Arena checkpoints just dropped.
🔍 Orionmist matches Lithiumflow but with Grounding/Search enabled; Lithiumflow is the base model.
🧪 Overall performance is better than ECPT, but still below earlier top checkpoints like X28/X58.
🎨 SVG panda and Pokeball visuals are strong with accurate color and solid lighting.
♟️ The chessboard makes good moves and beats ECPT in my testing.
🎮 The 3D Minecraft demo is smooth and similar to 2HT; X28 still wins on lighting.
🧰 The Blender Pokeball script works correctly with proper lighting setup.
🧠 General and math questions pass reliably, consistent with a solid base model.
⚖️ Responses feel more quantized/lower thinking budgets compared to earlier endpoints.
📉 Floor plan generation remains weaker, similar to ECPT results.
🔧 Expect Orionmist to handle recent events via Search; hoping for strong tool calling.
—
Timestamps:
00:00 – Introduction
00:34 – About Orionmist and Lithiumflow Gemini-3
02:05 – PhotoGenius (Sponsor)
03:22 – Testing & Results
09:09 – Ending
source