NaiveAI releases open weights for Naive-N0.5-Flash
NaiveAI released downloadable weights and inference code for Naive-N0.5-Flash on September 27 under an MIT license. The large model can be run on suitable local hardware, while a hosted API remains a later promise. CellCog independently confirms the release but its benchmark discussion relies on NaiveAI’s figures, so comparative performance remains unverified.
Artificial Intelligence··Midday
Weights and code are available
NaiveAI published the Naive-N0.5-Flash weights and inference code on Hugging Face on September 27 under an MIT license. CellCog separately recorded the release. That gives developers downloadable model files, although running them requires their own computing capacity. NaiveAI describes the system as an open-weight mixture-of-experts model: it has 309 billion parameters in total, with 15.5 billion active for each token. These are specifications from the developer’s model card, rather than independent measurements of performance.[1], [2]
Sparse attention shapes the long context
The model builds on Xiaomi’s MiMo-V2.5 base. NaiveAI says it changed the attention design, using 39 sliding-window layers and nine sparse-attention layers across a 48-layer network. A sliding window reads nearby tokens; a separate indexer scans the full history and sparse attention selects farther material. The company says the native context reaches one million tokens. Keeping the full key-value cache means sparse attention does not eliminate all memory costs. The model card lists about 315 gigabytes for FP8 weights before additional inference memory and calls for compatible NVIDIA graphics processors.[1]
Hosted access remains a future step
NaiveAI says it will provide API access later. CellCog reported that the endpoint was not yet live when it checked on September 27. The downloadable release therefore differs from an available hosted service. NaiveAI also publishes coding and AI research scores, but CellCog’s comparison tables draw on the company’s charts; they do not independently reproduce the tests. A team evaluating the model could inspect and run the released weights if it has suitable hardware, while the public record here establishes neither independent benchmark leadership nor a working API.[1], [2]
