minnesotanlp/Finch-9B

Finch-9B: Evolution Fine-Tuned for Discovery

Finch-9B is the largest and most capable model in the Finch family, developed by minnesotanlp. It is built upon the Qwen3.5-9B architecture and uniquely trained using Evolution Fine-Tuning (EFT). EFT is a novel mid-training approach that teaches LLMs how to evolve solutions by supervising them with evolutionary search trajectories from 355 diverse optimization tasks.

Key Capabilities

• Enhanced Discovery: Finch-9B acts as a powerful mutation operator within evolutionary search frameworks like OpenEvolve, significantly boosting the discovery of solutions for complex problems.
• Superior Performance: It outperforms its base model by up to +10.24% across 22 held-out tasks, with per-task gains reaching +290%. This includes strong results on NP-hard competitive programming problems (e.g., FrontierCS, CALICO's P263).
• Scalable Gains: The benefits of EFT scale with model size, making Finch-9B the top performer in its family for discovery tasks.
• Trajectory-Based Learning: The model learns from trajectories generated by a larger Qwen3.5-397B-A17B teacher model within the OpenEvolve scaffold.

Good For

• Evolutionary Search Systems: Ideal for integration into frameworks like OpenEvolve or SkyDiscover as a mutation operator to accelerate solution discovery.
• Complex Optimization: Applicable to a wide range of optimization problems where iterative improvement and novel solution generation are critical.
• Code Generation for Discovery: Demonstrates strong performance in generating improved programs for competitive programming and other computational tasks.

Limitations

Finch-9B's behavior is primarily validated with the OpenEvolve scaffold, and performance with other scaffolds is not guaranteed. It also inherits the characteristics and potential biases of its base model, Qwen3.5-9B.