Qwen3.6

// Overview

Qwen3.6 is the April 2026 release from the Qwen team at Alibaba Group, building on Qwen3.5. It is an open-weight line published under the Apache 2.0 license, distributed on Hugging Face and ModelScope and served through Alibaba Cloud Model Studio. The release leads with two open models: Qwen3.6-35B-A3B, a Mixture-of-Experts model with 35B total and roughly 3B active parameters (released 2026-04-16), and Qwen3.6-27B, a 27-billion-parameter dense model (released 2026-04-22) that the Qwen team frames as 'flagship-level coding in a 27B dense model.'

The headline story is parameter efficiency in agentic coding: the dense Qwen3.6-27B reaches 77.2 on SWE-bench Verified and 59.3 on Terminal-Bench 2.0, edging past the much larger Qwen3.5-397B-A17B MoE from the previous generation on most coding benchmarks. Qwen3.6 focuses on real-world utility — smoother front-end workflows, repository-level reasoning, and a 'thinking preservation' feature that keeps reasoning context across conversation history for iterative, agentic development.

Architecturally, Qwen3.6 uses a hybrid attention stack (Gated DeltaNet linear attention interleaved with periodic full Gated Attention) and Multi-Token Prediction for faster decoding. Both released models carry a vision encoder, accept text, image, and video input, and serve a native 262,144-token context that can be stretched toward ~1M tokens with YaRN RoPE scaling. As an Apache-2.0 release, the weights can be self-hosted via vLLM, SGLang, llama.cpp, MLX, or Transformers, or accessed through hosted APIs.

// Benchmarks

This model's scores

1. SWE-bench Verified77.2%
2. Terminal-Bench 2.059.3%
3. MMLU-Pro86.2%
4. AIME 202694.1%
5. GPQA Diamond87.8%
6. MMMU Pro75.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Qwen3.6-27B via OpenRouter (lowest listed provider); prices vary by host. The Apache-2.0 weights can also be self-hosted at no per-token cost.

Pricing source ↗

// Strengths

• Flagship-level agentic coding at small scale — the 27B dense model scores 77.2 on SWE-bench Verified, beating the prior-gen 397B MoE on most coding benchmarks
• Apache 2.0 open weights — free to self-host, fine-tune, and deploy commercially with no license gate
• Long context out of the box (262K native, extensible toward ~1M via YaRN scaling)
• Multimodal input — text, images, and video accepted via a built-in vision encoder
• Efficient hybrid architecture (Gated DeltaNet + periodic full attention) with Multi-Token Prediction for lower-latency decoding
• Thinking-preservation keeps reasoning context across turns, suited to iterative agentic workflows
• Broad ecosystem support: vLLM, SGLang, llama.cpp, MLX, Transformers, plus Qwen Code and Qwen-Agent

// Best for

• Agentic coding and repository-level software engineering (SWE-bench / Terminal-Bench style tasks)
• Front-end and full-stack web development workflows
• Self-hosted, privacy-sensitive deployments where Apache-2.0 weights matter
• Long-document and large-codebase analysis using the 262K-token context
• Multimodal tasks combining code, images, and video understanding
• Fine-tuning a strong open base for domain-specific coding or reasoning agents
• Cost-efficient inference where a 27B dense or 3B-active MoE beats running a much larger model

// How to access

// Qwen (open-weight) — every version

The full lineage of the Qwen (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.