ai/qwen3-coder-next-vllm repository overview

⁠Qwen3-Coder-Next

Qwen3-Coder-Next is an advanced open-weight language model designed specifically for coding agents and local development environments. This model represents a significant breakthrough in efficiency, achieving performance comparable to models with 10–20× more active parameters while using only 3B activated parameters from its 80B total parameter pool.

Built with a sophisticated Mixture-of-Experts (MoE) architecture featuring 512 experts with 10 activated per token, Qwen3-Coder-Next excels at long-horizon reasoning, complex tool usage, and recovery from execution failures. The model's 256k context length combined with its adaptability to various scaffold templates enables seamless integration with different CLI/IDE platforms including Claude Code, Qwen Code, Qoder, Kilo, Trae, and Cline, making it highly versatile for diverse development environments.

Through an elaborate training recipe, Qwen3-Coder-Next demonstrates robust performance in dynamic coding tasks and agentic workflows, making it ideal for sophisticated code generation, refactoring, debugging, and tool-calling scenarios. The model's cost-effective architecture ensures efficient deployment while maintaining high-quality output across various programming tasks.

⁠Characteristics

⁠Using this model with Docker Model Runner

docker model run qwen3-coder-next-vllm

For more information, check out the Docker Model Runner docs⁠.

⁠Architecture

The model features a hybrid architecture with the following components:

• Hidden Dimension: 2048
• Number of Layers: 48
• Hybrid Layout: 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))
• Gated Attention: 16 attention heads for Q and 2 for KV, head dimension of 256
• Gated DeltaNet: 32 linear attention heads for V and 16 for QK, head dimension of 128
• Mixture of Experts: 512 total experts, 10 activated per token, 1 shared expert, expert intermediate dimension of 512
• Non-Embedding Parameters: 79B

⁠Benchmarks

Qwen3-Coder-Next demonstrates exceptional performance across multiple coding benchmarks, achieving results comparable to models with significantly more activated parameters:

The model's efficiency allows for superior cost-effectiveness in agent deployment while maintaining competitive performance with larger models.

⁠Key Features

• Advanced Agentic Capabilities: Excels at long-horizon reasoning, complex tool usage, and recovery from execution failures
• Versatile IDE Integration: Compatible with multiple development platforms through its 256k context and scaffold template adaptability
• Efficient MoE Architecture: Uses only 3B activated parameters while achieving performance of much larger models
• Tool Calling: Native support for function calling and tool integration with custom tool templates
• Non-Thinking Mode: Optimized for direct code generation without explicit reasoning blocks

⁠Recommended Sampling Parameters

For optimal performance, use the following sampling parameters:

• Temperature: 1.0
• Top-p: 0.95
• Top-k: 40

⁠Links

• https://huggingface.co/Qwen/Qwen3-Coder-Next⁠
• https://qwen.ai/blog?id=qwen3-coder-next⁠
• https://github.com/QwenLM/Qwen3-Coder⁠
• https://qwen.readthedocs.io/en/latest/⁠

⁠Considerations

• The model requires significant GPU memory for full deployment (80B total parameters)
• For systems with memory constraints, consider reducing the context length from the native 262k to shorter values like 32,768 tokens
• Designed specifically for coding tasks; may not be optimal for general conversational use
• This model supports only non-thinking mode and does not generate <think></think> blocks in its output
• Best performance achieved with proper tool templates and scaffold integration in IDE environments

⁠Generated by

This model card was automatically generated using cagent-action⁠. Want to learn more about Docker Model Runner? Check out the project repository: https://github.com/docker/model-runner⁠.