Qwen 3.6 27B

Qwen 3.6 27B is the dense, natively multimodal flagship of Alibaba's Qwen3.6 family, released in April 2026. It directly builds on the dense Qwen3.5-27B, sharing the same hybrid attention design that mixes Gated DeltaNet linear attention with traditional full attention to cut KV-cache cost on long contexts. The model is a causal language model with an integrated vision encoder, processing text, image, and video inputs through both pre-training and post-training stages.

Compared with its predecessor, Qwen describes key gains in agentic coding, STEM reasoning, and vision tasks such as spatial intelligence, object localization, and document OCR. On the benchmarks Alibaba reports using its own internal agent scaffold, Qwen3.6-27B reaches 77.2% on SWE-bench Verified and 59.3 on Terminal-Bench 2.0, figures the team says exceed both Qwen3.5-27B and the much larger MoE [[sibling:qwen3-5-397b-a17b|Qwen 3.5 397B]] (397B total, 17B active).

A notable new feature is "Thinking Preservation," which retains reasoning traces across conversation history to reduce redundant token generation and improve KV-cache efficiency in multi-turn agent loops. Native context is 262,144 tokens.

Within the broader family, Qwen 3.6 27B sits alongside the sparse [[sibling:e2ee-qwen3-6-35b-a3b|Qwen 3.6 35B A3B FP8]] and [[sibling:qwen-3-6-plus|Qwen 3.6 Plus]], and the earlier [[sibling:qwen3-5-9b|Qwen 3.5 9B]] and [[sibling:qwen3-5-35b-a3b|Qwen 3.5 35B A3B]]. It ships as both BF16 and fine-grained FP8 checkpoints, supporting vLLM, SGLang, and Hugging Face Transformers.

Sources

Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model· qwen.aiQwen/Qwen3.6-27B · Hugging Face· huggingface.co

This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.