Qwen/Qwen3.5-2B-Base

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

Qwen/Qwen3.5-2B-Base 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 breakthroughs in multimodal learning and architectural efficiency, supporting a native context length of 262,144 tokens and extensible up to 1,010,000 tokens. This model excels in reasoning, coding, agents, and visual understanding, making it suitable for fine-tuning and research in multimodal AI applications.

Loading preview...

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

Qwen3.5-2B-Base is a 2.3 billion parameter causal language model developed by Qwen, designed for advanced multimodal AI applications. It features a unified vision-language foundation achieved through early fusion training on multimodal tokens, enabling cross-generational parity with Qwen3 and superior performance over Qwen3-VL models across various benchmarks.

Key Capabilities & Enhancements

  • Unified Vision-Language Foundation: Achieves strong performance in reasoning, coding, agents, and visual understanding by integrating vision and language from early training stages.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks combined with sparse Mixture-of-Experts for high-throughput inference with optimized latency and cost.
  • Scalable RL Generalization: Incorporates reinforcement learning scaled across million-agent environments for robust real-world adaptability.
  • Global Linguistic Coverage: Expanded support for 201 languages and dialects, facilitating inclusive worldwide deployment.
  • Extended Context Length: Supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.

Intended Use Cases

This model is primarily intended for fine-tuning, in-context learning experiments, and other research or development purposes. Its pre-trained nature and control tokens are optimized for efficient LoRA-style PEFT, making it a strong base for building specialized multimodal agents and applications.