NaiveAI releases Naive-N0.5-Flash: 309B MoE open-weight model under MIT

NaiveAI published Naive-N0.5-Flash on Hugging Face, an open-weight mixture-of-experts model with about 309 billion total parameters and 15.5 billion active, MIT-licensed weights and a native 1M-token context window. Its context handling combines Sliding-Window Attention and lightweight DeepSeek Sparse Attention in a roughly 5:1 layout with no full-attention layers.

• 309B total parameters, 15.5B active, MIT-licensed weights and code
• 1M-token context via hybrid SWA–DSA at roughly 5:1
• Built on the MiMo-V2.5 base with continued mid- and post-training
• NaiveRT serving stack claims up to 2,000 tokens/s in Ultrafast mode