Naive-N0.5-Flash

Naive-N0.5-Flash comes from NaiveAI, a start-up in Beijing, and targets coding and AI research work. It was published on 27 September 2026, and its weights and inference code are released under the MIT licence. NaiveAI's benchmark results are its own, and no independent evaluation had been published when the sources were read.

Features

• Coding work: a mixture-of-experts model meant for software development tasks.
• Research work: aimed at AI research tasks alongside coding.
• Long context: a native window of one million tokens, with sliding-window and sparse attention layers that keep long inputs cheaper to process.
• Scale: 309 billion parameters in total, with 15.5 billion active for each token.
• Open weights: the weights and inference code are published, so users can run the model themselves.

Who it works with

Developers and AI researchers who want an open model for coding and research work.

Key facts

• Developer: NaiveAI, a Beijing start-up.
• Released: 27 September 2026, with a research page and a Hugging Face repository.
• Base model: Xiaomi's open MiMo-V2.5, with further pre-training.
• Licence: MIT for the weights and inference code.