Alibaba’s Qwen3.6 series recently went open-source, including Qwen3.6-27B (dense) and Qwen3.6-35B-A3B (MoE). This update brings significant improvements in coding, context window, and architectural efficiency.
Model Specifications
| Spec | Qwen3.5-27B | Qwen3.6-27B | Qwen3.6-35B-A3B |
|---|---|---|---|
| Architecture | Dense | Dense | MoE (3B active) |
| Context | Default | 262K tokens | 262K tokens |
| Extended Context | — | Up to 1.01M | Up to 1.01M |
Qwen3.6-35B-A3B has 35B total parameters but only activates 3B during inference, making deployment costs close to a 3B model while achieving 30B+ level performance.
Benchmark Results
| Model | SWE-bench Verified |
|---|---|
| Qwen3.5-27B | 75.0 |
| Qwen3.5-397B-A17B | 76.2 |
| Qwen3.6-35B-A3B | Near Claude 4.5 Opus |
| Gemma4-31B | 52.x |
Qwen3.6-35B-A3B’s coding Agent capability approaches Claude 4.5 Opus — the first time an open-source model has come this close in this dimension.
Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-35B-A3B", torch_dtype="auto", device_map="auto")Hardware: Qwen3.6-35B-A3B runs on a single A100 40GB card thanks to MoE efficiency.
Main sources: