Naive AI open-sources Naive-N0.5-Flash, a 309B MoE model built with AI agents

Naive AI, the startup founded by Tsinghua associate professor Dai Jifeng, open-sourced its first model on September 28: Naive-N0.5-Flash, a 309B-total / 15.5B-active MoE with native 1M-token context, released under MIT on Hugging Face with inference code. Its pitch is “models built by models”: the official tech report says AI agents handled most architecture exploration, experiment execution and system optimization, while humans set goals and made the final calls.

The facts

• Specs: 309B total, 15.5B active parameters, post-trained on Xiaomi’s open-weights MiMo-V2.5 base; attention is a hybrid of 5 SWA layers and 1 DSA layer (DeepSeek Sparse Attention, top-2048 positions) with no full attention.
• License and pricing: MIT for weights and inference code; the API costs ¥0.6 per million input tokens, ¥2.6 output and ¥0.07 for cache reads.
• Self-reported benchmarks: PaperBench 63.2, MLE-bench-30 73.7%, NL2Repo 71.9, and a 73.8-second NanoGPT SpeedRun (previous record 77.5s) — all vendor-reported, independently unverified.
• AI-driven R&D: the NaiveRT optimization loop ran 151 experiments in 6 days and cut speculative-decoding latency from 12.3ms to 3.4ms; an AI-produced world model, AutoWM, is also claimed.
• Company: Naive AI was founded in February 2026 and is reportedly valued at $1.4 billion.

Our take

The story is the process, not any single score: if “AI runs experiments, humans make decisions” holds up and reproduces, the cost curve for mid-size labs chasing the frontier changes — the bottleneck shifts from headcount to compute budget. Two caveats: the model is post-training on an open base, not a from-scratch run, and the “built with AI” claims come from the company’s own report. The weights are on Hugging Face under MIT, so you can check directly; for a comparison point on the open-weights, agent-benchmark route, see Nex-N2.5.