Their open weight on device models are really impressive. Partly because I think they are the only ones out of all the frontier labs even working on local models.

> Gemma 4 31b is better for coding than Gemini

Is there a fine-tuned Gemma coding model? I'd assume that would perform quite well.