[Discussion] Open Source Model Selection

[MichaelClifford]

I wanted to start a discussion as well as start to keep a record of what open weight models we want to target for this project.

My initial proposals are:

• microsoft/Phi-4-mini-instruct
  Very small 3.8B model that has been instruction tuned, but not much if any code specific fine tuning.
• mistralai/Devstral-Small-2-24B-Instruct-2512
  Mid sized model that has been explicitly fined tuned for software engineering tasks

[TreeinRandomForest]

A few more:

Deepseek-Coder-V2

Qwen-2.5-Coder-32B

Codestral

We should also scan across #parameters e.g. Qwen has 3B and 7B models that set a nice baseline.