Qwen/Qwen3.6-35B-A3B

Hugging Face
VISIONConcurrency Cost:2Model Size:35BQuant:FP8Ctx Length:32kPublished:Apr 15, 2026License:apache-2.0Architecture:Transformer1.5K Open Weights Warm

Qwen3.6-35B-A3B is a 35 billion parameter causal language model with a vision encoder developed by Qwen, featuring 3 billion activated parameters and a native context length of 262,144 tokens. This model is specifically designed for agentic coding, excelling in frontend workflows and repository-level reasoning, and introduces a 'Thinking Preservation' feature to streamline iterative development. It demonstrates strong performance across various coding agent benchmarks and multimodal tasks, making it suitable for complex development and agent-based applications.

Loading preview...

Overview

Qwen3.6-35B-A3B is a 35 billion parameter causal language model with a vision encoder, developed by Qwen, building upon the Qwen3.5 series. It features 3 billion activated parameters and a native context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling. This release focuses on enhancing stability and practical utility, particularly for developers.

Key Capabilities

  • Agentic Coding: Significantly improved handling of frontend workflows and repository-level reasoning, making it more precise and fluent for coding tasks.
  • Thinking Preservation: A novel feature allowing the model to retain reasoning context from historical messages, which streamlines iterative development and can reduce token consumption in agent scenarios.
  • Multimodal Understanding: Supports both image and video inputs, demonstrating strong performance in STEM, puzzle, general VQA, text recognition, document understanding, and spatial intelligence benchmarks.
  • High Context Length: Natively supports 262,144 tokens, with extensibility to over 1 million tokens for ultra-long text processing.

Performance Highlights

Qwen3.6-35B-A3B shows competitive and often superior performance against models like Qwen3.5-27B and Gemma4-31B across various benchmarks:

  • Coding Agent: Achieves 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0, indicating strong capabilities in code generation and problem-solving.
  • Vision Language: Scores 81.7 on MMMU and 85.3 on RealWorldQA, showcasing robust multimodal reasoning.

Good for

  • Agent-based Development: Ideal for building sophisticated AI agents, especially those requiring complex coding, planning, and iterative reasoning.
  • Code Generation and Analysis: Excels in tasks involving frontend development, repository-level code understanding, and automated code fixes.
  • Multimodal Applications: Suitable for applications requiring understanding and processing of both text and visual (image and video) information, such as visual question answering and document analysis.
  • Long Context Processing: Beneficial for tasks that require processing extensive amounts of information, thanks to its large native and extensible context window.