Interesting. I wonder if there is opportunity to train a set of small model variants to excel at a certain stacks. Eg Qwen3.6-27B for Node + React or Qwen3.6-27B for Rust + TUI
Get comfortable with Deepseek's privacy policy for using this for anything serious.
"To improve and develop the Services and to train and improve our technology, such as our machine learning models and algorithms. Including by monitoring interactions and usage across your devices, analyzing how people are using it, and training and improving our technology."
> For Claude Sonnet 4.5, and GPT-5.1 we default to reporting high reasoning results, but when reported results are not available we use best available reasoning results.