Qwen3.8 is Alibaba Cloud’s latest flagship large language model series, introducing Qwen-Max-tier performance to the open-weights and developer ecosystem. Spanning variants from the locally runnable Qwen3.8-27B to the massive 2.4-trillion-parameter (95B active) Mixture-of-Experts (MoE) Qwen3.8-Max, the series delivers massive upgrades in long-horizon autonomous software engineering, native multimodal reasoning (text, image, and video), and customizable dynamic thinking modes.

The release of Qwen3.8 marks a major shift in frontier AI development, moving beyond single-prompt chatbots toward multi-day autonomous workflows and self-directed agents. Built on an expanded sparse Mixture-of-Experts architecture with native multimodal ingestion, Qwen3.8 is engineered to execute complex, multi-step engineering pipelines from scratch while drastically reducing inference drift.

Model Specifications & Key Tiers

Dimension / Model	Qwen3.8-27B	Qwen3.8-Max (2.4T)
Architecture	Dense, highly optimized for local hardware	Sparse Mixture-of-Experts (MoE, 95B active parameters)
Context Window	256K tokens (extendable to 1M via YaRN)	Up to 1,010,000 tokens (1M+ full context)
Hardware Target	Local consumer VRAM / unified memory (16GB–24GB+)	Cloud enterprise clusters & distributed quantization
Modalities	Text, Code, and Vision	Text, Code, High-Res Images, PDFs, and Extended Video
Key Strengths	Fast local debugging, low latency, agent integration	Multi-day autonomous coding, research replication, quant analysis

Core Breakthroughs & Capabilities

• Multi-Day Autonomous Coding: Unlike models designed purely for short code completions, Qwen3.8-Max is built for unattended engineering loops. In benchmark trials, it managed autonomous repositories over multiple days—creating issues, implementing features, running unit tests, and debugging CI failures end-to-end.

• Visual Ingestion as an Active Feedback Loop: Multimodality in Qwen3.8 is not just for input reading. In web and UI tasks, the model visually inspects its own generated frontend previews, catches misaligned elements or broken styling, and iterates automatically before returning the final build.

• Dynamic Thinking Control & Thinking Preservation: Developers can fine-tune reasoning depth using configurable parameters (none, low, medium, xhigh). The new Preserve Thinking feature retains previous reasoning traces across conversational turns, drastically reducing context loss during long debugging sessions.

• Broad Developer Ecosystem Compatibility: Qwen3.8 supports standard OpenAI- and Anthropic-compatible API endpoints, allowing immediate drop-in use across popular agent tools like Claude Code, Cursor, Cline, OpenCode, and local llama.cpp environments.