Robots Atlas>ROBOTS ATLAS
Qwen3.6-27B
AI Models›Qwen3

Qwen3.6-27B

3.6 · Family: Qwen3
Open-weight (Apache 2.0) 27B multimodal model from Alibaba's Qwen family. Hybrid Gated DeltaNet + Gated Attention, 262K-token context, strong at agentic coding.
✓ Active✓ Public access⚖ Open weightsMultimodalLLMReasoning model📁 Qwen3
Context window
262K
tokens
Parameters
27B
parameters
Release date
21 April 2026
🏢QwenProducer🏢AlibabaOwner
Access:DownloadHostedDeployment:💻 Local☁ Cloud

Overview

Qwen3.6-27B is an open-weight multimodal model released on 21 April 2026 by the Qwen team (Alibaba) under the Apache 2.0 license. It is the first open-weight variant of the Qwen3.6 generation, positioned around stability and real-world utility, with a particular focus on agentic coding (frontend workflows and repository-level reasoning).

The model pairs a language core with a vision encoder: it accepts text, image and video, and produces text and code. The language stack (27B parameters, 64 layers, hidden dimension 5120) uses a hybrid layout of Gated DeltaNet (linear attention) interleaved with Gated Attention at a 3:1 ratio, with grouped-query attention (GQA: 24 query heads, 4 KV heads). It is trained with Multi-Token Prediction (MTP).

The native context window is 262,144 tokens, extensible to roughly 1,010,000 tokens. A new “Thinking Preservation” option retains reasoning context from historical messages, streamlining iterative development. The weights are compatible with Hugging Face Transformers, vLLM, SGLang and KTransformers.

Classification
MultimodalLLMReasoning model
Family: Qwen3
Access & deployment
DownloadHosted
LocalCloud
Weights: Open weights
Key parameters
📏 Context: 262K
🧩 Parameters: 27B
✓ Tools · ✓ Fine-tuning
📥 Input: text, image, video

Technical specification

Context window
262K
tokens
Parameters
27B
parameters
License
Apache 2.0
Hardware requirements
Weights sharded across 15 safetensors files (bf16). Native context 262,144 tokens, extensible to ~1.01M tokens.
Features:✓ Tool use✓ Fine-tuning
Modalities
⬇ Input
textimagevideo
⬆ Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning — equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Agentic coding
Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).
Category: coding
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Long context
Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language
OCR
Recognising text within images and documents.
Category: vision

Benchmark results

11 benchmarks
SWE-bench
accuracy · SWE-bench Verified, agentic setting
77.2%
📄 technical_report
SWE-Bench Pro
accuracy · agentic setting
53.5%
📄 technical_report
Terminal-Bench 2.0
accuracy · agentic setting
59.3%
📄 technical_report
LiveCodeBench
accuracy · v6
83.9%
📄 technical_report
MMLU-Pro
accuracy
86.2%
📄 technical_report
GPQA
accuracy · GPQA Diamond
87.8%
📄 technical_report
AIME 2026
accuracy
94.1%
📄 technical_report
MMMU
accuracy
82.9%
📄 technical_report
MMMU-Pro
accuracy
75.8%
📄 technical_report
MathVista
accuracy · MathVista mini
87.4%
📄 technical_report
OCRBench
accuracy
89.4%
📄 technical_report

Technical architecture

Sources and related pages