Model testing

Model notes qwen3.8 - thinking off solves dark mode problem? qwen3.5:9b-mlx (thinking on?) - slow and failed, using pi qwen3.5:9b-mlx (thinking off) - fast and succeeded, using agent Qwen3.6-35B-A3B (unsloth │ ~17.7 │ MoE, 3B │ 73.4% SWE-bench Verified Nemotron-3.5-Lightning-30B-A3B │ 17.6 │ MoE, 3B gemma4:26b-mlx │ 18 GB │ same │ MLX runtime, Apple-native — free speed qwen3-coder-next~~ │ 52 GB │ 80B/3B Devstral Small 2 24B~~ │ 14.3 │ dense │ agent-first but dense 24B ≈ 10 tok/s here qwen3.6:27b~~ │ 17 GB │ dense │ high benchmarks, too slow interactively * Qwen3 Coder Next * Qwen3.5 27B * Qwen3.5 35B A3B * GLM 4.7 Flash * Devstral Small 2 24B