Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

A Tsinghua Professor's Naive.AI Ships a 309B Open-Source Coding Model, Built With AI in the Loop, After $400 Million in Three Rounds

A Tsinghua Professor's Naive.AI Ships a 309B Open-Source Coding Model, Built With AI in the Loop, After $400 Million in Three Rounds

Beijing-based Naive.AI released Naive-N0.5-Flash, a 309-billion-parameter mixture-of-experts model with 15.5 billion active parameters, one million tokens of context and an MIT licence, produced by post-training Xiaomi's MiMo-V2.5 rather than pre-training from scratch. The Information reports the startup has raised $400 million across three rounds at a $1.42 billion post-money valuation.

Naive.AI, a Chinese model company founded in February 2026 by a Tsinghua University associate professor, released its first public model on Sept. 29 in the form of Naive-N0.5-Flash. The company's legal name is Beijing Zhiyan Huisheng Technology, and the model is aimed squarely at coding and AI research workloads.

The technical specification is aggressive for an open release. Naive-N0.5-Flash has 309 billion total parameters in a mixture-of-experts design with 15.5 billion parameters active per token, natively supports one million tokens of context, and ships its weights and inference code under the permissive MIT licence. It is not trained from scratch: the model is built by post-training Xiaomi's open-source MiMo-V2.5. Of its 48 transformer layers, 39 use sliding-window attention and nine use DeepSeek-style sparse attention, and the network contains no full attention layers at all — an architecture choice the company says is aimed at long-document coding and research automation. A companion inference stack, NaiveRT, was itself optimized with AI in the loop; the company says it delivers 50 tokens per second per user in standard mode and up to 2,000 tokens per second in a turbo mode. Those throughput figures are company-reported and have not been independently verified.

The most unusual claim is about how the model was made. Naive.AI says AI models took part directly in the development process, writing code, running experiments, monitoring progress, analysing results and iterating. The company says it is also researching recursive self-improvement. Both are self-reported and currently unverifiable from outside.

The capital story is as striking as the model. The Information reports Naive.AI has completed three funding rounds this year totalling $400 million — $100 million, $180 million and $120 million — reaching a $1.42 billion post-money valuation, with Tencent, Sequoia China, IDG Capital and Jingwei Venture among the backers. That is one of the fastest capital accumulations recorded for a Chinese "Neo Lab" in seven months. The company has fewer than 100 employees.

Its founder, Dai Jifeng, took a bachelor's and a doctorate from Tsinghua's automation department before spending 2014 to 2019 as a principal researcher and research manager in the vision group at Microsoft Research Asia, then 2019 to 2022 as executive research director at SenseTime's research institute, returning to Tsinghua full time in 2022. The core team includes former MiroMind staff and specialists from SenseTime and MSRA. Until this week the company's website carried a single line — "100x Intelligence for the pioneers" — and it only opened an X account in September.

Naive.AI sits in a specific Chinese category. Alongside Pragmatik Labs, founded in August 2026 by former Alibaba Qwen head Lin Junyang, it is one of the country's two best-known Neo Lab projects, and both were valued near or above 10 billion yuan before shipping a product. That mirrors the pattern in Silicon Valley, where researchers left large labs to raise at headline valuations on the strength of a thesis and a résumé.

The strategy has an obvious trade-off. Post-training on an existing open-weight base is far cheaper than pre-training a frontier model from scratch, which is precisely why a team this small can ship a 309B model at all. But the ceiling is set by the base model, and MiMo-V2.5 is Xiaomi's, not Naive.AI's. The claim to watch is not the parameter count but whether developers adopt Naive-N0.5-Flash for real coding-agent work — long-context repository tasks where the mixed attention design and the million-token window are supposed to earn their keep. Until independent evaluations and API volume appear, the model's benchmark claims should be read as vendor-reported.

Comments (0)

Log in to join the discussion

Log In

No comments yet