Naive AI
Category: Foundation Models / LLMs
Beijing 'AI builds AI' lab that builds open-weight frontier LLMs for coding and AI R&D through AI-centered mid-training and post-training. Naive AI was founded in 2026. The company is led by Dai Jifeng (代季峰). Based in Beijing, China. Team size: 101-500. Total funding raised: $400M. Latest round: Venture Round. Key investors include Tencent, IDG Capital, Matrix Partners China (MPCi), HSG (HongShan, formerly Sequoia China).
- Founded
- 2026
- Headquarters
- Beijing, China
- Team size
- 101-500
- Total funding
- $400M
Value proposition
Uses AI agents to carry out most of the AI R&D loop: architecture exploration, training systems, inference optimization and experiments. The goal is to produce frontier-level open-weight models cheaply on top of open base models, with no from-scratch pre-training.
Products and solutions
Naive-N0.5-Flash: a 309B MoE model with 15.5B active parameters and native 1M context. It uses a hybrid SWA–DSA attention design with no full-attention layers and is MIT-licensed. NaiveRT: an AI-optimized inference runtime reaching up to about 2,000–2,122 tokens/s single-stream. API platform priced at $0.10 / $0.40 / $0.01 per million input, output and cache-read tokens. AutoWM: an AI-built world model that scored 77.43 on WorldArena-1 Track 1. AI-centered R&D sandbox infrastructure running about 10M sandboxes per week, with 100K concurrent at peak.
Unique value
A model built largely by AI and trained to do AI research itself, aimed at recursive self-improvement. It pairs open weights with ultra-fast inference.
Target customer
AI researchers, AI labs, developers and enterprises that need coding and AI research models, via open weights and a low-cost API
Industries served
AI research and development, software engineering and coding, developer tools
Technology advantage
Hybrid 5:1 sliding-window and DeepSeek Sparse Attention architecture with GQA4, trained on 3.25T tokens of continued training at 1M context. Mega-kernel fusion, programmatic dependent launch and speculative decoding give bitwise-deterministic fast decoding. Large-scale agent sandbox infrastructure. The founder has a strong track record in open-source vision and multimodal models.
How they differentiate
Naive AI does not pre-train from scratch. It starts from open base models and concentrates its compute on mid- and post-training, and it uses AI agents as the primary R&D workforce. It claims benchmark results above much larger models, such as 73.6 on SWE-bench Pro versus Qwen 3.8 Max, and leads on PaperBench and MLE-bench-30.
Main competitors
DeepSeek, Moonshot AI (Kimi), Zhipu AI (GLM), MiroMind
Key partnerships
Strategic investor Tencent
Major milestones
Feb 2026: founded in Beijing. Apr 2026: first reported funding of about $300M at about an $800M valuation. Sep 2026: total funding reached $400M at a $1.42B valuation. Sep 27, 2026: released the open-weight Naive-N0.5-Flash under the MIT license, along with NaiveRT.
Growth metrics
Raised about $400M across 3 rounds in 7 months, with a $1.42B post-money valuation as reported by The Information in Sep 2026. Fewer than 100 employees.
Market positioning
Well-funded Chinese frontier open-weight LLM newcomer, reported as a unicorn at a $1.42B post-money valuation after 7 months. It focuses on AI-for-AI and recursive self-improvement.
Geographic focus
China and global open-weight model ecosystem
Patents and IP
Open weights and inference code released under the MIT license (HuggingFace: NaiveAI/Naive-N0.5-Flash). There is a reported IP dispute with former affiliate MiroMind.
About Dai Jifeng (代季峰)
Associate Professor, Tsinghua University Dept. of Electronic Engineering (joined full-time July 2022); Ex-Microsoft Research Asia researcher; Ex-SenseTime senior research role; Ex-MiroMind (Shanda-incubated) co-initiator, 2025 to Jan 2026; Tsinghua University BS and PhD. Known for Deformable ConvNets (adopted as a standard PyTorch operator), BEVFormer and InternVL.
Latest news about Naive AI
More Foundation Models / LLMs companies
Official website: https://naive.ai