NaiveAI releases Naive-N0.5-Flash, an open model for coding and AI research

NaiveAI released its first model, Naive-N0.5-Flash, on September 27, 2026. According to the company's announcement, the model is designed for coding and for AI research and development work. Its weights are open on Hugging Face under the MIT license, so organizations with enough hardware can run it on their own servers.

What Naive-N0.5-Flash is

According to the model card, Naive-N0.5-Flash is a mixture-of-experts model with 309 billion parameters. In this design only part of the model runs at each step; here the active part is 15.5 billion parameters. The model can take in up to 1 million tokens at once and works with text only. The company did not train it from scratch: it built on Xiaomi's open-weight MiMo-V2.5 model, changed the attention layers and continued training.

How it stays fast on long inputs

Language models look back at the earlier text for every new piece they produce, and the work gets heavier as the text grows. According to NaiveAI, the model has no layer left that looks at the whole text at once. Of its 48 layers, 39 look only at a nearby window of 128 tokens. In the other 9, a small indexer scores the full history and the model looks only at the selected pieces. The company notes that the indexer still scans the full history.

AI models did the development work

The point the announcement puts forward is that AI models did much of the development work. In the company's account, the models write code, run experiments and analyze results, while researchers set the direction, constraints and criteria and make the critical decisions. The example given is NaiveRT, the inference software, which the company says was built in six days across 151 documented trials. Of those, 63 were adopted and 71 failed validation or were rolled back. When one line of work made results worse for seven rounds in a row, researchers stopped that direction. In most places the announcement does not name which models did this work.

How to run it

According to the model card, NVIDIA GPUs with FP8 support are required. The files of the compressed version take up about 315 GB, and more memory is needed on top of that to run it. The documented setup path is the Hugging Face Transformers library. No smaller version for personal computers has been released. The company also says it will offer API access; current prices are on the company's site.

By the company's measurement, NaiveRT reached more than 2,000 tokens per second for a single request on a system with eight GPUs. That is a peak value, measured over the best one-second window with thinking turned off. The announcement says NaiveRT is open source, and that its code will be shared by October 12. Neither the announcement nor the model card has a section on the model's limits. Before trying the model on your own work, it is worth following evaluations from independent researchers as well.

Sources: NaiveAI: Naive-N0.5-Flash announcement, Hugging Face: NaiveAI/Naive-N0.5-Flash, GitHub: NaiveAI-Labs/Naive-N0.5-Flash.