This model was retired on 30 October 2026 and no longer accepts API requests. Recommended replacement: DeepSeek V4 Flash 0731.
Qwen3.5 9B
Retired
The highly capable 9B-parameter multimodal model optimized for efficient agentic workflows and native tool calling.
Total Parameters
9B
Context Window
262K
Extensible to 1M+
Modalities
Text, Image
& Video
Function Calling
66.1%
BFCL-V4
Efficient Multimodal Intelligence at the 9B Scale
Qwen3.5 9B is a powerful foundation model that utilizes a hybrid Gated DeltaNet and Gated Attention architecture for highly efficient inference with reduced latency. Featuring 9 billion parameters, it delivers robust multimodal understanding across text, images, and video. Built with a native 262,144-token context window and explicit "Thinking Mode" capabilities, Qwen3.5 9B is engineered for production-grade reliability in autonomous agent workflows, advanced OCR, and complex global applications.
Hybrid Attention — 9B parameters
Built for agentic intelligence
Unified Multimodal Reasoning
Process text, video, and high-resolution images together. Excels at visual question answering, OCR document processing, and spatial reasoning.
Native Tool Calling & Agents
Production-ready function calling for multi-step agent orchestration, autonomous task planning, and reliable code generation.
Deep Reasoning ("Thinking Mode")
Built-in "Thinking Mode" generates explicit step-by-step reasoning traces before answering, dramatically increasing accuracy on complex tasks.
Global & Long-Context Processing
Analyze massive documents natively with the 262K context window (extensible to 1M+) while offering nuanced support across 201 languages.
Capable & Efficient
Proven performance across reasoning, coding, and agentic workflows for the 9B weight class.
Intelligence Index
Better than 75% of models
GPQA Diamond
Better than 83% of models
τ²-Bench Telecom
Better than 88% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | GPQA Diamond | 82% | Graduate-level scientific reasoning |
| Reasoning | Humanity's Last Exam | 13% | Humanity's Last Exam |
| Reasoning | τ²-Bench Telecom | 87% | AI agents in dual-control scenarios |
| Reasoning | AA-LCR | 59% | Long context reasoning evaluation |
| Reasoning | IFBench | 67% | Instruction-following accuracy |
| Reasoning | GDPval-AA | 20% | Agentic performance on real-world work tasks |
| Coding | SciCode | 29% | Python for scientific computing |
| Coding | Terminal-Bench Hard | 22% | Agentic coding & terminal use |
| Knowledge | AA-Omniscience Accuracy | 17% | Proportion of correctly answered questions |
| Knowledge | AA-Omniscience Non-Hallucination | 19% | Confidently answered questions that are correct |
Metrics sourced from Artificial Analysis.