On the evening of April 22, 2026, Alibaba's Qwen team quietly dropped a bombshell on Hugging Face: Qwen3.6-27B, a 27B dense model, is now open-source under Apache 2.0. It scored 59.3 on Terminal-Bench 2.0, tying Anthropic's closed flagship Claude 4.5 Opus. On SWE-bench Verified, it hit 77.2, beating the previous 397B MoE flagship (76.2) while shrinking model size from 807GB to 55.6GB — a 14x reduction. SkillsBench jumped from 27.2 to 48.2, a 77% gain.
Specs & Deployment: 16.8GB Quantized Version Runs on 24GB GPUs
Qwen3.6-27B is a pure dense architecture, activating all 27B parameters per inference. Native context length is 262,144 tokens, extendable to ~1M via YaRN. The full model weighs 55.6GB; Q4_K_M quantization brings it down to 16.8GB, allowing direct loading on consumer GPUs with 24GB VRAM or Mac M-series hardware. License is Apache 2.0, free for commercial use. Recommended deployment: SGLang ≥0.5.10 or vLLM ≥0.19.0. The model also integrates a vision encoder for image, text, and video understanding.
The Local Brain for OpenClaw and Hermes Agent
OpenClaw, a popular local agent app akin to Claude Code, was blocked by Anthropic in early April. The team then urged users to switch to local models like Qwen. Qwen3.6-27B directly fills this gap. Hermes Agent, led by NousResearch, emphasizes a self-learning skill loop. Both frameworks suffered from the lack of a truly capable local model. Qwen3.6-27B scored 72.4 average and 60.6 Pass^3 on Claw-Eval (a coding agent benchmark), surpassing Claude 4.5 Opus's 59.6, offering a credible local option.
Alibaba's Open-Source Strategy and the Shifting US-China AI Landscape
Qwen3.6-27B follows the open-source release of Qwen3.6-35B-A3B (MoE) on April 16. The dense variant fills the gap for fully local deployment, while closed-source flagships like Qwen3.6-Plus remain monetized via API. This open-close strategy builds ecosystem trust while cashing in on cutting-edge models. The broader context: Meta is reportedly abandoning its open-source AI push and turning to Alibaba's Qwen to train a paid model called Avocado. US tech giants are pulling back from open-source; Chinese firms are doubling down. The question for developers is no longer "to open or not to open," but "which open model is good enough." Qwen3.6-27B offers the clearest answer yet.