Dai Jifeng's Startup Naive AI Unveils First Open-Source Model Naive-N0.5-Flash
On September 28, 2026, Naive AI, the startup founded by Dai Jifeng, released its first open-source model, Naive-N0.5-Flash. The model has 309B total parameters and 15.5B activated parameters, natively supports a 1M-token context, and is designed for programming and AI research and development.
Founded in February 2026, Naive AI has been exploring an AI-native research workflow in which AI models handle coding, experiment execution, process monitoring and result analysis, while human researchers focus on setting goals, constraints and evaluation criteria and making final decisions. The technical report released with the model says AI participated in exploring attention architectures and optimizing training, inference and deployment systems.
In one internal case, researchers and AI completed 151 rounds of optimization experiments in six days, cutting the end-to-end time of one speculative decoding step in the NaiveRT inference runtime from 12.3 ms to 3.4 ms. In a world model research task, Naive-N0.5-Flash ran 15 major experiments over about 400 hours to produce AutoWM, which scored 77.43 on WorldArena-1 Track 1 video quality evaluation, above the previously published best score. The model also scored 63.2 on PaperBench and 73.7% on MLE-bench-30.
The model is built on the open-weight MiMo-V2.5 base and adopts a hybrid architecture combining sliding window attention and DeepSeek sparse attention. Model weights and inference code are released under the MIT license, and an API will be provided.
