Naive-N0.5-Flash, the model NaiveAI says it built with AI's help

On 27 September 2026, Beijing-based startup NaiveAI published Naive-N0.5-Flash on Hugging Face under an MIT license. As the model card explains, this is not an architecture built from scratch but an adaptation of Xiaomi's open-weight MiMo-V2.5 base model, further trained on an additional 3.25 trillion tokens. The system uses a Mixture-of-Experts design with 309 billion total parameters, 15.5 billion of them active per token, and handles a native context of one million tokens by combining sliding-window attention layers with DeepSeek Sparse Attention. It is not entirely clear, however, how the MIT license squares with the terms of Xiaomi's base-model license.

What sets the project apart is how it was built: the company's technical blog describes it as AI-centred research and development. According to the blog, human researchers set the direction, constraints and criteria and make the critical decisions, while implementation, optimisation, experimentation and analysis are delegated to AI models. The same text states that “human researchers set the direction, define constraints and criteria, and make critical decisions. The human edge lies in experience, intuition and judgement.” The independent outlet Runtime Wire, however, noted that the published documents do not make clear how much work the AI actually did, how that contribution was measured, or whether the approach really lowers future development costs. The same outlet links the company to Jifeng Dai, an associate professor at Tsinghua University who previously worked at Microsoft Research Asia and SenseTime.

On performance and cost, the model card lists API prices of $0.10 per million input tokens and $0.40 per million output tokens, while Runtime Wire reports the price list in yuan: 0.60 yuan for input and 2.60 yuan for output per million. The two figures come from different sources and do not match the exchange rate exactly. For inference, which requires Nvidia GPUs with FP8 support, the company offers its proprietary NaiveRT system, credited with a standard speed of 50 tokens per second per user, rising to 2,000 in Ultrafast mode. The company blog places that peak on an eight-GPU setup without naming the GPU model, while Runtime Wire specifies that the measurement of more than 2,100 tokens per second refers to a decoding test limited to a single stream.

All the performance benchmarks shown in the official GitHub repository come from an internal evaluation protocol called AutoResearch, built on Claude Code. Since the results are presented only as charts and have not been checked by independent third parties, the exact scores claimed by the startup cannot be confirmed. The lack of outside confirmation calls for caution — a recurring theme whenever release speed outpaces the verifiability of the metrics.

If handing code-writing to another machine speeds up releases, the real asymmetry lies in verification: validating an AI's work still takes the human judgement the startup claims as its own safeguard. Until the benchmarks are measured with something other than an internal protocol and replicated by others, the claimed performance should be read as the company's claims, not as verified data. — Olya