the local qwen models are very impressive, even the small 1.7b param qwen3 model doesn’t seem stupid like the gemma ones do from my experience.