Alibaba's Qwen3.6-27B Achieves Flagship-Level Coding on Single Consumer GPU

Alibaba's Qwen team released Qwen3.6-27B on April 22, 2026, a 27-billion-parameter dense model under Apache 2.0 that outperforms the previous-generation Qwen3.5-397B-A17B across key agentic coding benchmarks. Community testing shows it hitting 80 tokens per second on a single RTX 5090 GPU with a 218,000-token context window. On SWE-bench Verified, Qwen3.6-27B scores 77.2, up from 75.0 for Qwen3.5-27B and competitive with Claude 4.5 Opus at 80.9. On Terminal-Bench 2.0, it matches Claude 4.5 Opus exactly at 59.3. On QwenWebBench, an internal bilingual front-end code generation benchmark, it scores 1487, a 39% jump from 1068 for Qwen3.5-27B. On SWE-bench Pro, it reaches 53.5, beating Qwen3.5-27B at 51.2 and the much larger Qwen3.5-397B-A17B at 50.9.

Qwen3.6-27B introduces Thinking Preservation, a mechanism that retains reasoning traces across conversation history, reducing redundant token generation and improving KV cache efficiency in multi-turn agent workflows. It uses a hybrid architecture blending Gated DeltaNet linear attention with traditional self-attention. The model supports both multimodal thinking and non-thinking modes. Full weights are available on Hugging Face under Apache 2.0 with no commercial restrictions. Community testing on Reddit's LocalLLaMA reports approximately 80 tokens per second on a single RTX 5090 using vLLM 0.19 with MTP enabled and a 218k context window. Other configurations hit 45 tokens per second in LM Studio and 1,157 tokens per second on dual 4090s with 16 concurrent requests. The model fits on 18GB VRAM in quantized form, making frontier-level coding assistance available to individual developers without enterprise GPU clusters or API costs.