The open-weight LLM landscape sees a significant upgrade in reasoning capabilities with the release of G9v3-39A5B. Moving beyond the G9v3-3B version, AI9Stars has implemented a Mixture of Experts (MoE) architecture featuring 39 billion parameters and 5 active experts, specifically engineered for high-level logical processing.

Benchmarking the Intelligence Leap

According to data from Artificial Analysis, G9v3-39A5B achieves a score of 31 on the Intelligence Index. This represents a substantial jump from the 16.1 recorded for the 3B model and places it well above the median score of 9 for similar open-weight models. This performance boost is particularly evident in coding tasks and complex tool-use workflows.

Dual Reasoning Modes: Think vs No Think

G9v3-39A5B introduces a versatile operational framework supporting both Think and No Think modes. This allows users to toggle between deep, iterative reasoning—essential for complex problem solving—and streamlined, direct responses for standard assistant tasks, providing a level of control typically reserved for larger proprietary systems.

Open Licensing and Deployment

Released under the Apache 2.0 license, G9v3-39A5B is fully open for both personal and commercial applications. With weights available on Hugging Face and the repository hosted on GitHub, the model is positioned as a powerful alternative for developers seeking high-reasoning capabilities within local or private cloud environments.