NaiveAI releases a 309B model built with AI-assisted research

Why it matters

NaiveAI is selling both an open-weight model and a method for developing future models with AI doing more of the research work. Its speed and capability claims remain company-reported, making the model's actual performance and the practical contribution of AI to its development the key tests.

NaiveAI released Naive-N0.5-Flash on September 27th, an open-weight model for coding and AI research that the Beijing startup says was developed with AI systems handling substantial parts of the research and engineering work. The launch post on X describes a 309-billion-parameter mixture-of-experts model with 15.5 billion parameters active per token, a native 1-million-token context window and an inference runtime called NaiveRT.

The model's release puts a product behind NaiveAI's central proposition: AI systems can help build the next generation of AI. In its technical launch write-up, NaiveAI says models wrote code, ran experiments, monitored results and proposed iterations during development. Human researchers set objectives, constraints and evaluation standards, and made key decisions. The company presents this as a research process it intends to extend, with Naive-N0.5-Flash itself trained for AI research tasks.

That division of labor matters to the claim. NaiveAI is describing AI-assisted research under human direction, rather than a model independently designing and shipping its successor. The write-up says the model's architecture and systems work emerged from that process; it does not establish how much work was performed by AI, how the company measured that contribution or whether the approach lowers the cost of developing later models.

NaiveAI is associated with Jifeng Dai, a Tsinghua University associate professor whose career spans academic research and senior industry roles. Dai's Tsinghua profile lists a bachelor's degree and Ph.D. from the university, a principal research manager role at Microsoft Research Asia from 2014 to 2019, and executive research director at SenseTime Research from 2019 to 2022. The Information has reported that Dai leads NaiveAI. His profile credits him with work spanning computer vision and multimodal models, including InternVL.

The architecture makes a specific trade: keep a very long context while avoiding full-attention layers. NaiveAI says the model combines sliding-window attention with DeepSeek Sparse Attention, selecting a subset of earlier tokens for some attention calculations. In the company's explanation, the design reduces attention computation and memory traffic at long context lengths; the sparse indexer still scans the history, and the model still retains its full key-value cache. The company says the model starts from Xiaomi's MiMo-V2.5 base and acknowledges Xiaomi's MiMo team and DeepSeek's sparse-attention work.

NaiveAI's headline speed figure needs its test conditions attached. The X post says NaiveRT can reach 2,000 tokens per second in Ultrafast mode. The longer write-up reports a peak of 2,122 tokens per second on eight GPUs in a single-stream test, measured over the best one-second window from 41 HTML and SVG generation requests. The reported figure excludes prompt processing, so it describes a narrow decode-speed result rather than an end-to-end experience on arbitrary workloads. NaiveAI also compares a full speculative-decoding cycle of 3.4 milliseconds with 12.3 milliseconds for SGLang on the same system; these are company-reported measurements.

The release says the model weights and inference code are available under the MIT license. NaiveAI's write-up lists API prices of ¥0.60 per million input tokens, ¥2.60 per million output tokens and ¥0.07 per million cached tokens. Those rates give the release a commercial route alongside downloadable weights, while the long-context coding and research pitch competes for workloads where users may care about both cost and the amount of task history a model can process.

https://x.com/naiveailab/status/2104247060186951725

Naive-N0.5-Flash's benchmark positioning remains the company's own claim. The launch post calls it top-tier in coding and leading in AI research and development; the write-up describes internal evaluations and example tasks. The numbers explain the model's design and the company's performance case, but they do not by themselves establish superiority over other open models. NaiveAI is asking the market to assess two propositions together: that its model can handle demanding coding and research work, and that AI can make frontier-model development more automated.