Qwen ✓

Handles complex software projects more autonomously, producing complete deliverables across extended development workflows.

Improves professional work across research, data analysis, writing, legal, financial, and other specialized tasks.

Maintains stronger performance on long-horizon goals with less human intervention from initial planning through final execution.

Combines text, image, and video understanding with tool use for more capable multimodal agent workflows.

Scales to 2.4 trillion parameters while becoming the first Max-class release with openly available model weights.

Stronger coding across frontend prototyping, debugging, multi-file engineering, and production-grade kernel optimization.

Sustains coherent autonomous execution across multi-hour workflows involving hundreds or thousands of tool calls.

Improves office automation through tool integrations, multi-agent orchestration, data analysis, document creation, and spreadsheet work.

Generalizes more reliably across different agent frameworks without depending on harness-specific shortcuts.

Handles long-term planning with greater resistance to context degradation, instruction drift, and failed intermediate steps.

Stronger agentic coding across frontend work, terminal operations, automated task execution, and repository-level problem solving.

Better long-horizon agent performance through tighter integration of reasoning, memory, and tool execution, with gains on planning and tool-calling benchmarks.

Sharper multimodal reasoning for document understanding, visual analysis, video reasoning, visual coding, and fine-grained grounding.

Larger default context handling with a 1M context window plus stronger ultra-long-context information extraction.

More stable real-world agent workflows with broader assistant compatibility and a preserve_thinking API option to maintain prior reasoning context on multistep tasks.

Native multimodal VLM: open-weights release of Qwen3.5-397B-A17B, positioned as a vision-language model with strong reasoning, coding, agent capability, and multimodal understanding.

397B total, 17B active: sparse MoE design activates 17B parameters per forward pass out of 397B, targeting high capability with lower runtime cost.

Hybrid attention for inference efficiency: fuses linear attention (Gated Delta Networks) with sparse MoE to optimize speed and cost without dropping performance.

Massive multilingual expansion: language and dialect coverage increases from 119 to 201, aiming for broader global usability.

Qwen3.5-Plus hosted tier: 1M context by default with built-in tools and adaptive tool use, plus modes like Auto (tool use), Thinking (deep reasoning), Fast (instant) for different latency and cost needs.

Incorporates hybrid reasoning capabilities for improved adaptability and efficiency.

Available in both dense and sparse model architectures.

Trained on 36 trillion tokens across 119 languages and dialects.

Supports a 128K token context window for larger models.

Enhanced reasoning capabilities with modes for both complex logical tasks and general-purpose chat.

Introduced over 100 open-source models, including specialized versions for coding, mathematics, and multimodal tasks.

Enhanced capabilities in various domains, outperforming previous versions.

Introduced both dense and sparse model architectures.

Some models were open-sourced, while others remained proprietary.

Enhanced reasoning capabilities and performance benchmarks

Introduced as a beta version of Qwen 2.

Offered in multiple sizes: 0.5B, 1.8B, 4B, 7B, 14B, 32B, and 72B parameters.

Included a 14B Mixture-of-Experts (MoE) model with 2.7B active parameters.

Enhanced multilingual support and a stable 32K context length across all sizes.

Significant improvements in chat performance and human preference alignment.