G9v3-39A5B

G9v3-39A5B is an open-weights Mixture-of-Experts (MoE) causal language model developed by the AI9Stars team. Designed with a focus on strong reasoning capabilities, the model contains approximately 39 billion total parameters while activating only 5 billion parameters per token during inference. This architecture is intended to balance high-level cognitive performance with the efficiency typical of smaller dense models, making it a practical choice for local and self-hosted environments.

The model is optimized for a range of complex tasks including coding, multi-step reasoning, and tool-use integration. It features a 131,072-token context window, allowing it to process and analyze large documents or extensive conversation histories. Performance benchmarks suggest that G9v3-39A5B delivers competitive intelligence in logic-heavy and agentic workflows, often achieving parity with larger models while maintaining lower computational overhead.

A notable feature of the G9v3-39A5B is its support for both Think and No Think modes within a single checkpoint. This capability allows the model to toggle its internal reasoning process, which can help in reducing hallucinations and improving the accuracy of structured outputs. The model is released under the Apache 2.0 license, supporting broad application in both personal and commercial projects.