Qwen 3.6 27B
27 billion parameters, dense architecture - outperforms its 397B MoE predecessor
Qwen 3.6 27B is a dense model built on the Hybrid Gated DeltaNet architecture with 64 layers and 262K native context. It scores 77.2% on SWE-bench Verified, surpassing the 397B MoE Qwen 3 at 76.2%, while fitting in ~55.6GB VRAM at FP16 or ~18GB with quantization.
Model variants
Dense architecture, maximum quality per parameter
Qwen 3.6 27B delivers frontier-class performance in a dense 27B form factor. Choose the instruction-tuned variant for chat and agentic tasks, or the base model for fine-tuning.
Hybrid Gated DeltaNet Architecture
27B dense parameters, 64 layers, hidden dimension 5120
Qwen 3.6 27B uses a Hybrid Gated DeltaNet design that combines linear attention efficiency with gated recurrence. The 262K native context window is extensible to 1M tokens, making it ideal for long-document analysis and complex agentic workflows.
With qwen 3.6 27b vram requirements of ~55.6GB at FP16 or ~18GB quantized, this model runs on a single high-end GPU or dual consumer GPUs with the qwen 3.6 27b gguf format.
Instruction-tuned
27B Instruct
Optimized for conversational AI, coding, and complex agentic tasks
Fine-tuned for instruction following, multi-turn dialogue, and tool use via the qwen 3.6 27b api
Pre-trained
27B Base
Foundation dense model for fine-tuning and specialized applications
Pre-trained on diverse data with Hybrid Gated DeltaNet architecture for maximum quality
Capabilities
Dense powerhouse that punches above its weight class
Qwen 3.6 27B combines the Hybrid Gated DeltaNet architecture with 262K context to deliver performance that surpasses models 14x its size on real-world coding benchmarks.
Elite software engineering
77.2% on SWE-bench Verified - beating the 397B MoE Qwen 3 (76.2%). The qwen 3.6 27b benchmark results prove dense architectures can match frontier-scale models on real-world coding.
Terminal mastery
59.3 on Terminal-Bench 2.0, matching Claude 4.5 Opus. Handles complex multi-step terminal workflows, debugging sessions, and system administration tasks with expert-level proficiency.
Advanced reasoning
94.1% on AIME 2026 mathematics and 86.2 on MMLU-Pro knowledge reasoning. Step-by-step thinking mode enables transparent problem solving across math, logic, and science.
262K to 1M context
262K native context window extensible to 1M tokens. Process entire codebases, long research papers, and multi-turn conversations without losing coherence.
Competitive coding
83.9 on LiveCodeBench v6 for competitive programming. Excels at algorithmic problem solving, code generation, and complex debugging tasks.
Practical skill execution
48.2 on SkillsBench, surpassing Claude 4.5 Opus (45.3). Demonstrates superior ability to follow complex instructions and execute multi-step real-world tasks.
Key highlights
Exceptional qwen 3.6 27b benchmark results
Qwen 3.6 27B achieves frontier-class results across coding, reasoning, and agentic benchmarks while maintaining efficient dense inference.
Top achievements
- SWE-bench Verified: 77.2% - beats 397B MoE predecessor (76.2%)
- Terminal-Bench 2.0: 59.3 - matches Claude 4.5 Opus
- SkillsBench: 48.2 - beats Claude 4.5 Opus (45.3)
- AIME 2026: 94.1% mathematics
- LiveCodeBench v6: 83.9 competitive coding
Technical specs
- 27B dense parameters, 64 layers, hidden dimension 5120
- Hybrid Gated DeltaNet architecture
- 262K native context, extensible to 1M tokens
- qwen 3.6 27b vram: ~55.6GB FP16, ~18GB quantized
- Available in qwen 3.6 27b gguf format for local deployment
Performance
Dense 27B that outperforms 397B MoE on real-world coding
Qwen 3.6 27B scores 77.2% on SWE-bench Verified and 94.1% on AIME 2026, proving that a well-architected dense model can match or exceed models many times its size.
The qwen 3.6 27b benchmark suite demonstrates consistent excellence across software engineering, terminal operations, mathematics, and competitive coding - rivaling or surpassing models with 10x+ more parameters.
SWE-bench Verified: 77.2% - surpasses 397B MoE Qwen 3 (76.2%)
Terminal-Bench 2.0: 59.3 - matches Claude 4.5 Opus
SkillsBench: 48.2 - beats Claude 4.5 Opus at 45.3
AIME 2026: 94.1% on advanced mathematics
MMLU-Pro: 86.2 across diverse knowledge domains
Benchmark comparison
Qwen 3.6 27B vs frontier models
Qwen 3.6 27B delivers frontier-class performance across software engineering, terminal operations, reasoning, and coding benchmarks. Access results via the qwen 3.6 27b api.
| Benchmark | Qwen 3.6 27B Dense Featured | Qwen 3 235B A22B MoE | Claude 4.5 Opus Proprietary | Qwen 3.6 35B A3B MoE |
|---|---|---|---|---|
SWE-bench Verified Real-world software engineering | 77.2% | 76.2% | - | 73.4% |
Terminal-Bench 2.0 Terminal operations | 59.3 | - | 59.3 | 51.5 |
SkillsBench Real-world task execution | 48.2 | - | 45.3 | - |
AIME 2026 Mathematics No tools | 94.1% | - | - | 92.7% |
LiveCodeBench v6 Competitive coding | 83.9 | - | - | 80.4 |
MMLU-Pro Knowledge & reasoning | 86.2 | - | - | - |
Benchmark results from official Qwen 3.6 model card and HuggingFace evaluations.
Hybrid Gated DeltaNet
A new architecture that redefines dense model efficiency
The Hybrid Gated DeltaNet architecture combines linear attention with gated recurrence across 64 layers and a hidden dimension of 5120. This design enables 262K native context extensible to 1M tokens while maintaining the inference simplicity of a dense model.
- 64 layers with hidden dimension 5120 for deep representation learning
- 262K native context window, extensible to 1M tokens
- qwen 3.6 27b vram: ~55.6GB FP16, ~18GB with quantization (qwen 3.6 27b gguf)
Software Engineering
77.2% SWE-bench Verified - the dense model that beat a 397B MoE
Qwen 3.6 27B achieves 77.2% on SWE-bench Verified, surpassing its 397B MoE predecessor at 76.2%. Combined with 59.3 on Terminal-Bench 2.0 (matching Claude 4.5 Opus) and 83.9 on LiveCodeBench v6, it's a complete software engineering assistant accessible through the qwen 3.6 27b api.
- 77.2% SWE-bench Verified - real-world GitHub issue resolution
- 59.3 Terminal-Bench 2.0 - expert-level terminal operations
- 83.9 LiveCodeBench v6 - competitive programming excellence
Get started
Try Qwen 3.6 27B now
Start chatting instantly via the qwen 3.6 27b api, or download weights for self-hosted deployment.
Local deployment
Run Qwen 3.6 27B on your hardware
Deploy locally with qwen 3.6 27b gguf quantized weights. Runs on consumer hardware with ~18GB VRAM.
Qwen ecosystem
Part of the Qwen 3.6 model family
Qwen 3.6 27B is part of Alibaba's latest model family, with dense and MoE variants, extensive community support, and broad framework compatibility.
Get started
Ready to build with Qwen 3.6 27B?
Start chatting instantly for free, or download qwen3.6-27b weights for self-hosted deployment on your infrastructure.