Hacker News front pageJun 30, 2026, 01:05 UTC+8

Qwen 3.6 27B is the sweet spot for local development

Piotr Migdał tested Qwen 3.6 27B and calls it the first local model that works as a general intelligence. Running 8-bit quantized on a Macbook Max M5 128GB with llama.cpp and multi-token prediction, it hits 32 tok/s using 42GB RAM. It handled constrained writing, generated a hexagonal minesweeper npm package in one shot, and built a reactive landing page. The post includes full llama.cpp setup commands and recommends against Ollama on ethical grounds.

Why it matters: A first-person experiment with real numbers, not a press release. Qwen 3.6 is already a hot topic, and this piece adds practical local-deployment details. Score capped at 72 because it's a personal review, not an official launch or major product update.

Read the original ↗Export Markdown

Take

This post is worth clicking because it's a real-world local model experience, not a benchmark table. Piotr Migdał ran Qwen 3.6 27B at 8-bit quantization on a Macbook Max M5 with 128GB RAM, using llama.cpp with multi-token prediction. He got 32 tok/s at 42GB memory usage. It handled constrained writing, generated a hexagonal minesweeper npm package in one shot, and built a reactive landing page — all from single prompts.

I'd discount this a bit: the tasks lean creative and frontend, not complex backend logic or long debugging chains. But 32 tok/s on a laptop is genuinely usable, and he included the full llama.cpp command so you can copy-paste it. He avoids Ollama on ethical grounds, which he's written about elsewhere.

If you have a 128GB M5 Max, this is worth trying. If you don't, the 42GB RAM footprint means most machines can't run the 8-bit version — you'd need to go lower on quantization, and quality will drop.

Source: Hacker News front page · quesma.com
Published: Jun 30, 2026, 01:05 UTC+8 · Jun 30, 2026