Qwen3.5-9B

Qwen3.5-9B is a 9-billion-parameter open-weight vision-language model from Alibaba's Qwen Team, released in March 2026 under the Apache 2.0 license. It uses a hybrid architecture combining Gated Delta Networks with Gated Attention in an 8×(3×DeltaNet→FFN→1×Attention→FFN) layout, plus a vision encoder for native multimodal understanding. The model supports 262,144 tokens natively (extendable to ~1M via YaRN/RoPE scaling), covers 201 languages, and includes a thinking mode for explicit chain-of-thought reasoning.

On official benchmarks, Qwen3.5-9B scores 82.5% on MMLU-Pro, 91.1% on MMLU-Redux, 88.2% on C-Eval, 81.7% on GPQA Diamond, and 91.5% on IFEval. For coding and agents it reaches 65.6% on LiveCodeBench v6, 66.1% on BFCL-V4, and 79.1% on TAU2-Bench. On vision tasks it achieves 78.4% MMMU, 89.2% OCRBench, and 84.5% VideoMME. Independent evaluations report ~82.93% HumanEval pass@1 and ~89.5% GSM8K accuracy.

Community reception is strong for its parameter class: it ranks #2 among tracked image-text models on HuggingFace with 1,800+ likes, and is widely regarded as one of the best 9B-class open models for coding, RAG, and agent workloads. It significantly outperforms Qwen3.5-4B and prior-generation 8B models while remaining deployable on consumer GPUs via quantization.