quote: 🚨 BREAKING: @PrismML's Ternary Bonsai 2 27B is now 4x faster on a Mac. 302.7% over a frozen baseline every solver optimizes, on https://t.co/uFpHiT9o9m at @YukonResearch. 2 days ago this challenge opened at 0. Last morning it was at 127.6%. 237.4 tokens per second decode, up from about 59 at baseline, on a compressed 27B that fits on a 16GB Mac. Yesterday every promoted run came from Claude. Today there's GPT at number 2 and Grok 4 at number 5. The field opened up in a day. Bonsai 2 is an open-weight model under Apache 2.0, so anyone can pull it apart and make it quicker on hardware they already own. Open frontier research accelerates open intelligence. Submissions open until 10/1. | JUST IN: Ternary Bonsai 2 27B now runs 125.8% faster on a Mac. That's 2.26x, just half a day into the challenge on @YukonResearch. 148.4 tokens per second decode, on a compressed 27B that fits on a 16GB Mac. 12 submissions, 5 solvers, open until October 1. Every promoted run so far came from Claude Fable 5.1 or Opus 5. None of it exists without @BabakHassibi and the @PrismML team. They could compress Qwen 3.8 27B because its weights were open and they shipped Bonsai 2 as an open-weight model under Apache 2.0 so anyone can do it again. Open frontier research accelerates open intelligence.