unsloth/Qwen3.5-2B

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Feb 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3.5-2B is a 2.3 billion parameter causal language model developed by Qwen, featuring a unified vision-language foundation and an efficient hybrid architecture. It integrates multimodal learning and architectural efficiencies, supporting a native context length of 262,144 tokens. This model excels in multimodal reasoning, coding, and agentic tasks, offering expanded linguistic coverage across 201 languages and dialects.

Loading preview...

Qwen3.5-2B: A Multimodal Agent Foundation Model

Qwen3.5-2B is a 2.3 billion parameter causal language model developed by Qwen, designed for prototyping, task-specific fine-tuning, and research. It features a unified vision-language foundation, enabling early fusion training on multimodal tokens. This allows it to achieve cross-generational parity with Qwen3 and outperform Qwen3-VL models across various benchmarks, including reasoning, coding, agentic tasks, and visual understanding.

Key Capabilities & Features

  • Unified Vision-Language Foundation: Seamlessly processes and understands both visual and linguistic inputs.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks combined with sparse Mixture-of-Experts for high-throughput inference with minimal latency.
  • Scalable RL Generalization: Incorporates reinforcement learning scaled across millions of agent environments for robust real-world adaptability.
  • Global Linguistic Coverage: Supports 201 languages and dialects, enhancing its applicability worldwide.
  • Long Context Window: Features a native context length of 262,144 tokens, ideal for complex, long-form tasks.
  • Agentic Usage: Excels in tool calling capabilities, with recommended integration via Qwen-Agent and Qwen Code.

Performance Highlights

Qwen3.5-2B demonstrates strong performance across various benchmarks. In language tasks, it scores 66.5 on MMLU-Pro (Thinking mode) and 78.6 on IFEval (Thinking mode). For vision-language tasks, it achieves 64.2 on MMMU and 76.7 on Mathvista (mini), showcasing its robust multimodal reasoning. The model supports both non-thinking and thinking modes, with specific sampling parameters recommended for optimal results in different task types.

Should you use this for your use case?

This model is particularly well-suited for applications requiring strong multimodal understanding, long context processing, and agentic capabilities. Its efficient architecture makes it suitable for scenarios where high-throughput inference is critical. Developers focusing on multilingual applications or those needing a model capable of complex reasoning and tool use will find Qwen3.5-2B a valuable asset.