Naive AI Releases N0.5-Flash Model for Automated AI Research
The AMW Read
A new research-focused model and MIT release meaningfully update the foundation-model player map, while its broader impact depends on independent evidence of reliable research gains.
Naive AI Releases N0.5-Flash Model for Automated AI Research
Naive AI has released Naive-N0.5-Flash, an open-weight model designed to read research papers, reproduce experiments and support model development. According to Leiphone, the company began with an existing open-source base, replaced its global attention layers with a mix of sliding-window and sparse attention, and trained the resulting model on 3.25 trillion tokens across three stages. The company says the model uses 15.5 billion active parameters and scored 73.6 on SWE-bench Pro. Leiphone also reports that the weights and inference code carry an MIT license, with API access priced at RMB 0.6 per million input tokens and RMB 2.6 per million output tokens.
The release puts Naive AI on the foundation-model map with a specific bet: improve an existing base through architecture changes and continued training, then use the resulting model in the research process itself. That approach targets a costly bottleneck for model labs: the many short experiments needed to test training methods, long-context behavior and inference efficiency. The reported benchmark comparisons and claims of autonomous research capability are company claims in this account; they do not establish how reliably the system performs across independent research projects.
For builders, the MIT release makes the model and inference code available to test against real paper-reproduction and experiment-management workflows. For investors, the useful evidence will be repeatable research output and measured savings in researcher time or compute, beyond a single benchmark score. Those results would show whether Naive AI's research-focused model can become a durable part of model development rather than another capable coding model.