Qwen 3.6 27B is the sweet spot for local development
Points and comments are a snapshot, not live.
Qwen 3.6 27B is the best local model for coding on a Macbook Max M5.
Qwen 3.6 comes in two variants: a fast 35B MoE and a slower, more powerful dense 27B model. The 27B fits in 28 GB RAM at 8-bit, runs at 32 tokens/second on an M5 with MTP, and creates working applications from single prompts (e.g., hexagonal minesweeper, candle shop landing page). The author recommends llama.cpp over Ollama for running it locally, and provides specific quantization and launch commands. They place it at roughly mid-2025 frontier model quality, better than Gemma 4 31B.
What commenters are saying
Most commenters confirm the model works well locally, with several sharing their setups: one runs it on a 3090 at 50-70 t/s, another on an M4 Pro 24GB. There's agreement that 27B at Q4 or IQ4_NL fits in 32GB RAM or less. Some find the 35B MoE variant fast enough for interactive use. Multiple users report Gemma 4 12B struggles with coding tasks by comparison. One commenter notes Ornith 1.0 (based on Qwen 3.5) as another strong option worth trying.