Qwen 3.6 27B

Qwen 3.6 27B is a local-first, dense multimodal LLM purpose-built for code generation, tool use, and complex agent workflows. It is not a generic chatbot but a 'local coding brain,' outperforming even some 198B-class models in tests. Free and open-weight, it runs best on a DGX Spark or high-end consumer GPU; NVIDIA's NVFP4 quantization reaches 110 tok/s at 256k context. Stock 27B needs quantization for practical use. Community demos estimate electricity at ~$9/month, far cheaper than cloud APIs. The main hurdle is technical setup with llama.cpp and quantized model downloads. Ideal for developers and tinkerers wanting offline, private, cost-free inference; not suitable for users without discrete GPUs or those needing simple chat. It is a customizable inference engine, not a turnkey app. Evidence comes from X hands-on benchmarks (NVFP4 vs UD-Q8, Tetris coding, concurrency) and a Zhihu deep-dive into quantization merges. Quality is high, with strong consensus on its top-tier local code/agent performance; the only common caveat is high power draw and heat output.