Qwen 3.6 27B makes local-first coding genuinely usable

A developer who had written off local language models reports that Qwen 3.6 has changed his mind. The dense 27B variant — slower but stronger than its mixture-of-experts sibling, the 35B A3B — handled real tasks from single prompts, including scaffolding a working hexagonal Minesweeper as a proper pnpm package where the faster MoE model cut corners. The takeaway echoes the prevailing Hacker News sentiment that the model ‘punches above its weight’: you trade speed for output quality that holds up against frontier APIs on everyday coding work.

The practical pitch is that running it is now trivial. The author recommends llama.cpp over Ollama (the latter on stated ethical grounds), pulling an 8-bit quantization with multi-token prediction from Hugging Face and pointing an agent like OpenCode at the local server. Performance is the crux: roughly 30 tokens/sec on a MacBook M5, with a commenter hitting 50 tokens/sec on an RTX 5090 at heavier quantization. Eight-bit quantization barely dents quality, while more aggressive 2–4 bit schemes (as used by rivals like a quantized DeepSeek V4 Flash) trade away more.

The broader argument is about ownership. Local models can be fine-tuned, run offline, kept away from US or Chinese cloud providers for sensitive or medical data, and — unlike proprietary frontier models that get pulled or run on unsustainable subsidies — can’t be taken away. The piece frames Qwen 3.6 as a stepping stone toward open-weight models like GLM 5.2 that approach frontier quality, and bets that future architectures will offload factual knowledge to tool calls, shrinking the weights enough to run serious intelligence on consumer hardware.