Qwen 3.6-27B Open-Source: Dense Model Matches 397B MoE on Single GPU
Key Questions
How does Qwen 3.6-27B compare to much larger models?
The 27.8B dense Qwen 3.6-27B model matches the performance of a prior 397B MoE model, scoring 77.2% on SWE-bench. It runs affordably on a single RTX 3080 at roughly $7 per hour.
Is Qwen 3.6-27B suitable for local agentic workflows?
Yes, its Apache-2.0 license, strong benchmark results, and single-GPU capability make it practical for local high-quality agent deployments and cost-sensitive applications.
What quantized releases support local Qwen deployment?
Quantized Qwen3.5 checkpoints co-designed with inference engines have been released to enable efficient local running on consumer hardware.
Alibaba released Qwen 3.6-27B, a dense 27.8B model under Apache-2.0 that matches previous 397B MoE performance while running on a single RTX 3080. SWE-bench score 77.2%, cost ~$7/hr. Game-changer for local agentic workflows and affordable high-quality open models. Additionally, Qwen3.5 quantized checkpoints have been released, co-designed with inference engines, further enabling local deployment. An article highlights the broader shift from model strength to deployment cost, with Alibaba's Tongyi Qianwen (Qwen) as a key example of open-source edge applications.