According to CoinMeta, NaiveAI, founded by Dai Jifeng, has released naive-n0.5-flash and made the model weights available for public access. This model is based on modifications to Xiaomi's mimo-v2.5 and base, and it is open-sourced using the MIT protocol. NaiveAI retains the original MOE architecture but focuses on improving the attention mechanism. The team replaced the global attention mechanism with a sliding-window attention system and DeepSeek sparse attention, and continued to train the model with 3.25 trillion tokens. The modified model natively supports 1 million token contexts, which reduces the computational load when processing extremely long texts. AI was directly involved in the development, responsible for writing code, running experiments, analyzing results, and making further optimizations, while the researchers were in charge of setting the direction and making key decisions. Taking the inference system Naivert as an example, the team conducted 151 rounds of optimization experiments within 6 days, of which 63 rounds were adopted. The official announced highest inference speed is 2122 tok per second, but the testing conditions were quite special: 8 GPU were used, the thinking mode was disabled, the input processing time was not calculated, and the speed was measured from the best performance in 41 requests over one second. In standard mode, the official reported a speed of about 50 tok per second per user.