Term
Qwen 3.6
Qwen 3.6 is Alibabas open-weights model from April 2026 — the dense 27B variant matches much larger MoE models on agentic coding benchmarks, under an Apache 2.0 license.
Qwen 3.6 — explained in more detail
Qwen 3.6 is Alibabas open-weights model family; the dense Qwen3.6-27B variant was released on April 22, 2026. As the first dense (non-MoE) model of the 3.6 line it ships under the permissive Apache 2.0 license, so it can be freely downloaded, self-hosted and used commercially. Notably, the 27B variant surpasses larger Mixture-of-Experts models such as the previous generations 397B-A17B model on select agentic coding benchmarks.
Key facts
- Release: April 22, 2026, access via open weights (BF16 and FP8-quantized).
- Architecture: dense 27B-parameter model, 64 layers mixing Gated DeltaNet linear attention with classic self-attention.
- Context window: native 262,144 tokens, up to about 1,010,000 tokens via YaRN scaling.
- License: Apache 2.0.
- Benchmarks: SWE-bench Verified 77.2, SWE-bench Pro 53.5, Terminal-Bench 2.0 59.3 (level with Claude Opus 4.5), GPQA Diamond 87.8, AIME26 94.1, LiveCodeBench v6 83.9.
Example / Practical use
Qwen 3.6 targets operators who want to self-host a strong, freely adaptable model — for coding agents, data analysis or multi-step agentic workflows. Its “Thinking Preservation” retains reasoning traces across the conversation history, reducing redundant token generation in multi-turn scenarios. The dense architecture keeps the model competitive on coding tasks despite a comparatively small parameter count.
Delimitation
Within the Qwen family, 3.6-27B is the dense entry variant; above it sit larger MoE models and the newer Qwen 3.8 line. Against proprietary top models such as Claude Opus or GPT Sol a gap remains on the hardest tasks, but the open, Apache-licensed access and self-hosting capability are the decisive differences. Direct comparison models are DeepSeek V3.1 and other open weights.