Alpie Seed 9B returns a decision in 261ms. Its rivals take 0.4 to 0.8 seconds.

One game in every lane: same serve, same rules, same question, asked of a different model in each. One answer moves the ball one step, so every ball travels at its own model's answer time and nothing is skipped or sped up. Alpie Seed is a 9B model from 169Pi with its thinking switched off, racing models of about its size or larger, where the lab publishes a size, that also answer without thinking first, each from a different lab: open models on free keys with no credit card, and Gemini 3.5 Flash-Lite and Nova 2 Lite on Google's and Amazon's own APIs, billed per call. Pong is the Atari game chat models handle worst, so a loop that won't wait is the fair place to see what a fast answer is worth.

Gemini 3.5 Flash-Lite races on a billed key: Google's free tier allows it 15 requests a minute and a lane here makes about 60. Nova 2 Lite is billed per call on Amazon Bedrock; Amazon does not publish its size. Gemma 4 and Alpie Core 32B, 169Pi's reasoning model, raced here until 26 October 2026 and are now measured in the comparison instead.

Inspired by Ably Labs' open-source Jev Pong. The how-it-works page puts the same question to more models, including both Alpie models with thinking on and Grok 4.7, which cannot stop reasoning. See the numbers.