QwenLM/Qwen3.8

What it solves

Qwen3.8 is a series of open-weight large language models designed to handle complex, multi-step tasks with higher reliability. It specifically targets improvements in coding, professional research, and long-horizon agentic tasks, bringing "Max-class" performance to the open-source community.

How it works

Built on the architectural foundation of Qwen3.5, the series evolves through several iterations:

• Qwen3.8: Focuses on autonomous planning and environment feedback to improve end-to-end task completion. It introduces reasoning_effort to tune reasoning depth and preserve_thinking to retain reasoning context across messages.
• Qwen3.6: Prioritizes stability and real-world utility, specifically enhancing agentic coding for front-end workflows and repository-level reasoning.
• Qwen3.5: Uses a unified vision-language foundation with early fusion training, a hybrid architecture (Gated Delta Networks and sparse Mixture-of-Experts), and reinforcement learning scaled across million-agent environments.

Who it’s for

Developers and enterprises looking for high-performance open models for software development, autonomous agents, and multimodal applications across 201 languages and dialects.

Highlights

• Agentic Capabilities: Stronger autonomous planning and better handling of environment feedback for reliable task completion.
• ** uma Unified Multimodal Foundation**: Early fusion training on trillions of tokens for parity across reasoning, coding, and visual understanding.
• Flexible Thinking Control: Ability to tune reasoning depth and preserve thinking context in conversations.
• Broad Compatibility: Supported by major inference engines like vLLM, SGLang, and TokenSpeed, and training frameworks like Unsloth and Llama-Factory.