NaiveAI's Naive-N0.5-Flash hits 2,000 tokens/sec for coding

Naive-N0.5-Flash did not start from scratch: NaiveAI built it by continuing to train Xiaomi's MiMo-V2.5, a sparse Mixture-of-Experts model with 310B parameters and 15B active parameters trained on 48T tokens. The team adapted those weights to DeepSeek Sparse Attention layers through multi-stage training that took roughly 3.25T additional tokens on top of the original 48T, with about 3T of those at a fixed 1M context length and a constant learning rate.

The model's pitch extends beyond speed. NaiveAI describes it as largely built by AI and trained to do AI research itself, with recursive self-improvement as the long-term goal. In practice, the trade-off found is that Naive-N0.5-Flash takes longer to reason through harder problems, yet still delivers answers faster overall, at a slightly higher cost per question.

On app-building tests, Naive-N0.5-Flash was more consistent than GLM 5.3 Flash (high) and held its own against DeepSeek V4.1 Flash 1017, producing the most detailed solar system app and coming a close second on ray-tracer efficiency. A caution is that the speed sample was small, at 10 questions per difficulty level, and answers were not graded for accuracy.