Naive-N0.5-Flash: 309B MoE model combining 1M context with 2,000 tokens/s speed

This is an open-weight MoE model with 309B parameters and 15.5B active parameters. It is specialized for coding and AI R&D tasks, with weights and inference code released under the MIT license.

It adopts a hybrid structure combining Sliding-Window Attention and DeepSeek Sparse Attention. It supports native 1M token context without full attention layers, making it advantageous for processing large codebases or long documents.

Through its proprietary inference system NaiveRT, it offers generation speeds of 50 tokens per second in standard mode and up to 2,000 tokens in Ultrafast mode. AI-centric optimization techniques such as Mega-kernel fusion and speculative decoding are applied.

API pricing is set at $0.10 for input, $0.40 for output, and $0.01 for cache reads (per million tokens). It balances cost and performance for development environments requiring both long context retention and high inference speed.