Alibaba's Qwen3.6 Shrinks to Fit a Single Consumer GPU With 2.5M Downloads
A community AWQ INT4 quantization of Qwen3.6-27B shrinks Alibaba's flagship-tier coding model to about 20GB, making it self-hostable on a single consumer GPU.
Takeaways
- A community AWQ INT4 quant of Qwen3.6-27B has crossed 2.5M downloads on Hugging Face.
- The 20.44 GB checkpoint runs on a single RTX 4090 or 5090 with ~17 GB VRAM.
- Base model scores 77.2% on SWE-bench Verified and matches Claude 4.5 Opus on Terminal-Bench 2.0.
- Hybrid Gated DeltaNet plus Gated Attention gives 262K native context, extensible to 1M with YaRN.
- Ships multimodal with text, image, and video inputs under Apache 2.0.
- Introduces Thinking Preservation to keep reasoning traces across multi-turn agent loops.
A community-built 4-bit quantization of Alibaba's Qwen3.6-27B has crossed 2.5 million pulls on Hugging Face, turning what was a data-center-class coding model into something you can run on a single high-end consumer GPU. The cyankiwi/Qwen3.6-27B-AWQ-INT4 repository packages the full multimodal Qwen3.6-27B into a 20.44 GB compressed-tensors checkpoint using Activation-aware Weight Quantization (AWQ), calibrated on a STEM and agentic dataset. It ships under the same Apache 2.0 license as the base model.
Flagship coding scores in a self-hostable package
Qwen3.6-27B is Alibaba's open-weight dense language model with 27 billion parameters, support for text, image, and video inputs, and a native 262,144-token context window extensible to 1,010,000 tokens. It is also the first open-source model to introduce Thinking Preservation, which retains chain-of-thought reasoning traces across multi-turn conversations.
On a series of agentic coding benchmarks, the 27B dense model edges past the much larger 397-billion-parameter Qwen3.5 mixture-of-experts model from earlier this year, and matches Claude 4.5 Opus on Terminal-Bench 2.0. On SWE-bench Verified, Qwen3.6-27B scores 77.2% versus Claude Opus 4.6's 80.8%, and on Terminal-Bench 2.0 it scores 59.3%, matching Claude 4.5 Opus exactly.