← Back to feed

Alibaba’s Qwen team released Qwen3.6-27B on April 22, a dense 27-billion-parameter model under Apache 2.0 that beats the previous generation’s 397-billion-parameter flagship on every major coding benchmark while reducing the model file from 807 GB to 55.6 GB. A Q4_K_M quantized build fits in 16.8 GB — a single high-end consumer GPU.

On SWE-bench Verified, Qwen3.6-27B scores 77.2%. The Qwen3.5-397B-A17B scored 76.2%. The gap is one point, but the direction matters: the smaller model wins. More telling is Terminal-Bench 2.0, where Qwen3.6-27B hits 59.3% against the 397B model’s 52.5%. On SkillsBench the margin is wider still: 48.2 versus 30.0.

The 77.2% SWE-bench Verified score ties GPT-5.4, placing a 27B open-weight model running locally at the same level as one of OpenAI’s frontier proprietary offerings on this benchmark.

Architecture: Where the Efficiency Comes From

The gains trace to a Gated DeltaNet hybrid architecture. Each of the model’s 64 layers follows a repeating pattern: three Gated DeltaNet linear-attention blocks followed by one standard self-attention block. Linear attention scales with context length rather than quadratically, so memory pressure at long contexts drops sharply. The single standard attention layer per group anchors precise retrieval without giving up the efficiency of the linear layers.

This 3:1 ratio (linear to standard attention) is the same structural choice Qwen made in the 3.6-35B-A3B MoE model released five days earlier. The dense 27B carries it into the larger parameter regime without the routing complexity of a mixture-of-experts design.

Native context is 262,144 tokens, extensible to approximately 1 million tokens via YaRN. Inference compatibility: SGLang 0.5.10+, vLLM 0.19.0+, KTransformers, Hugging Face Transformers.

Thinking Preservation

The second major addition is Thinking Preservation. In agent workflows, models typically regenerate their reasoning chain from scratch on each turn. Qwen3.6-27B retains the chain-of-thought across conversation turns, reducing redundant token generation and improving KV cache efficiency in multi-turn sessions. This is a practical win for agentic coding work where long-running sessions accumulate significant context.

A single checkpoint handles both thinking and non-thinking modes. The preserve_thinking option enables the retention behaviour without switching models or variants.

The Benchmark Spread

Benchmark	Qwen3.6-27B	Qwen3.5-397B	Qwen3.6-35B-A3B
SWE-bench Verified	77.2%	76.2%	73.4%
SWE-bench Pro	53.5%	50.9%	n/p
Terminal-Bench 2.0	59.3%	52.5%	51.5%
SkillsBench	48.2	30.0	n/p
QwenWebBench	1487	n/p	n/p

QwenWebBench covers bilingual front-end code generation across web design, apps, games, SVG, data visualization, animation, and 3D. The 1487 score is a 39% jump from the Qwen3.5-27B’s 1068.

The Sizing Argument

The practical point is not merely that the 27B model beats the 397B on benchmarks. It is that this performance is now accessible on hardware that most developers own.

At BF16, Qwen3.6-27B requires 55.6 GB — a pair of 3090s, or a single A100. At Q4_K_M quantization, the footprint drops to 16.8 GB. That fits on a single RTX 4090 or RTX 5090, with community benchmarks showing approximately 80 tokens per second on the 5090.

The Qwen3.5-397B-A17B at 807 GB requires dedicated server infrastructure. The model it just outscored on coding benchmarks runs on a gaming GPU.

Context

This is the second model in the Qwen3.6 family. The first, Qwen3.6-35B-A3B, is a sparse MoE with 35B total parameters and 3B active, released April 16. It scored 73.4% on SWE-bench Verified. The new 27B dense model scores 3.8 points higher on the same benchmark with a different architectural approach.

Alibaba also maintains closed Qwen3.6 variants (Qwen3.6-Plus, Qwen3.6-Max-Preview) available through API. The open-weight releases serve a different audience: local deployment, fine-tuning, and cost-sensitive production inference.

The Apache 2.0 licence means no restrictions on commercial use, modification, or distribution.