NaiveAI's 1000 Tokens/s Challenge to DeepSeek

🔥 What happened Chinese startup NaiveAI launched Naive-N0.5-Flash, an open-weights coding model built on Xiaomi's MiMo-V2.5 and retrained with 3.25T extra tokens at 1M context. It hits nearly 1,000 tokens per second and beats DeepSeek V4.1 Flash on NL2Repo-Bench (71.9 vs 68.2). 💡 Why it matters Median response time is 3.1 seconds versus DeepSeek's 10.1. On hard HLE questions with thinking enabled, it averages 108,500 completion tokens — nine times GLM 5.3 Flash. The catch: on hard problems it's actually slower than DeepSeek (89.8 vs 79.7 seconds), and the authors didn't grade accuracy. ⚡ Our take Recursive self-improvement as a pitch smells like marketing — but 1,000 TPS on coding tasks is a real lever. Betting on speed over reasoning depth is a gamble. Until someone grades the answers, this is a benchmark beauty pageant.