Beijing startup NaiveAI open-sources 309B MoE Naive-N0.5-Flash under MIT
NaiveAI, a Beijing startup led by Tsinghua’s Jifeng Dai and reportedly valued at $1.4B, released Naive-N0.5-Flash on September 27: 309B-total / 15.5B-active MoE with 1M-token context and a hybrid SWA + DeepSeek sparse attention architecture.
Aggregated coverage on AI Weekly on September 28 and earlier reports on Pandaily and StartupFortune on September 27 say Beijing-based startup NaiveAI, founded by Tsinghua’s Jifeng Dai and reportedly valued at around $1.4 billion, released Naive-N0.5-Flash as an MIT-licensed open-weights mixture-of-experts model targeting coding and AI R&D. The model has 309 billion total / about 15.5 billion active parameters with a native 1M-token context window.
Naive-N0.5-Flash is built on Xiaomi MiMo-V2.5 as the base, stacked with 39 sliding-window attention layers and 9 lightweight DeepSeek Sparse Attention layers, with no full-attention layers; total training tokens are 3.25 trillion. NaiveAI says the company’s R&D and engineering pipelines were substantially executed by AI systems, with humans setting objectives and evaluation standards. NaiveAI also reports an ultrafast mode of up to 2,000 tokens per second on its in-house NaiveRT inference stack, with API pricing of $0.10 per million input tokens and $0.40 per million output tokens.
NaiveAI separately disclosed more than 7,000 RL training environments in its pipeline, framing the Naive family as a representative sample of "AI-led R&D." Naive-N0.5-Flash is NaiveAI’s first open-weights flagship release, and Hugging Face co-founder and CEO Clément Delangue reshared third-party evaluations the same day. NaiveAI’s prior round was led by Sequoia China and valued the company at about $1.4 billion.