RedHatAI/Qwen3.6-35B-A3B

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3.6-35B-A3B is a 35 billion parameter Mixture-of-Experts (MoE) causal language model with a vision encoder, developed by Qwen. This model is optimized for agentic coding, handling frontend workflows and repository-level reasoning with enhanced fluency and precision. It features a native context length of 262,144 tokens, extensible up to 1,010,000 tokens, and supports multimodal inputs including text, images, and video.

Loading preview...

Overview

Qwen3.6-35B-A3B is the first open-weight variant of the Qwen3.6 series, a 35 billion parameter Mixture-of-Experts (MoE) causal language model with a vision encoder. It is designed to provide a more intuitive and productive coding experience, prioritizing stability and real-world utility. The model has 3 billion activated parameters, making it efficient for hardware with more compute than memory bandwidth. It supports a native context length of 262,144 tokens, which can be extended up to 1,010,000 tokens using YaRN scaling techniques.

Key Capabilities

  • Agentic Coding: Enhanced capabilities for frontend workflows and repository-level reasoning.
  • Thinking Preservation: New option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.
  • Multimodal Input: Supports text, image, and video inputs, making it suitable for diverse applications.
  • High Context Length: Natively handles 262,144 tokens, extensible to over 1 million tokens.
  • Performance: Achieves strong results across various benchmarks, including SWE-bench Verified (73.4), Terminal-Bench 2.0 (51.5), and QwenWebBench (1397), demonstrating proficiency in coding and agentic tasks.

Good For

  • Developers: Ideal for those building agentic applications, especially for coding tasks, due to its specialized agentic coding and thinking preservation features.
  • Resource-Constrained Environments: Its MoE architecture with 3 billion activated parameters makes it suitable for deployment on hardware with higher compute but limited memory bandwidth.
  • Complex Multimodal Tasks: Excels in scenarios requiring understanding and generation based on text, images, and video, as evidenced by its strong performance on MMMU (81.7) and RealWorldQA (85.3) benchmarks.
  • Long-Context Applications: Beneficial for tasks requiring extensive context, such as analyzing large codebases or lengthy documents, with its support for up to 1,010,000 tokens.