Qwen3.5-9B

Qwen3.5-9B is the largest model in the Qwen 3.5 Small Series and arguably the most impressive for what it hits relative to its size. At 9 billion parameters, it beats the previous generation's Qwen3-30B on most language benchmarks and dominates GPT-5-Nano and Gemini-2.5-Flash-Lite on vision tasks by double-digit margins.

The 9B strikes a rare sweet spot: it's small enough to run on consumer hardware but capable enough to compete with models 3-9x its size on serious benchmarks. At BF16, it fits on a single RTX 3090 or any 24GB GPU. With 4-bit quantization, it drops to roughly 5GB - viable on a RTX 3060 12GB or a M1 Mac with room to spare.

Key Specifications

The 9B uses the deeper 32-layer stack (shared with the 4B) and the widest hidden dimension (4,096) among the small models. Its FFN intermediate dimension of 12,288 gives it meaningful capacity for world knowledge and reasoning. The vision encoder is a DeepStack Vision Transformer with Conv3d patch embeddings for temporal video understanding.

Benchmark Performance

Language (Thinking Mode)

The 9B beats the previous Qwen3-30B on GPQA Diamond by 8 points, IFEval by 3 points, and LongBench v2 by 10 points. It even edges past Qwen3-80B on GPQA Diamond and IFEval - models that required an A100 cluster to run.

Vision-Language

The vision results are the 9B's strongest selling point. It beats GPT-5-Nano by 13 points on MMMU-Pro, 17 points on MathVision, and 32 points on document understanding. Against the 27B (3x its size), it trails by only 4 points on MMMU and 7 points on MathVision - a narrow gap for a 3x reduction in compute.

Key Capabilities

Native multimodal at every level - Unlike Qwen 3 which needed separate VL model variants, the 9B handles text, images, and video from the same weights. The DeepStack Vision Transformer uses Conv3d patch embeddings to capture temporal dynamics in video, and merges features from multiple encoder layers rather than just the final one. VideoMME at 84.5 with subtitles shows genuine video comprehension, not just frame sampling.

Long context without the cost - The Gated DeltaNet linear attention maintains constant memory complexity, enabling 262K native context with extension to roughly 1M tokens via YaRN. LongBench v2 at 55.2 and AA-LCR at 63.0 confirm the model actually uses this context effectively, not just theoretically supports it. For comparison, the previous Qwen3-80B scored 48.0 on LongBench v2 despite being nearly 9x larger.

Thinking and non-thinking modes - The model supports both explicit chain-of-thought reasoning (enclosed in <think>...</think> tags) and direct response generation. Thinking mode is enabled by default and can be toggled via the enable_thinking API parameter or /think and /no_think tags. Thinking mode pushes GPQA Diamond to 81.7; non-thinking mode still delivers competitive results for latency-sensitive applications.

Pricing and Availability

Apache 2.0 licensed and available on HuggingFace and ModelScope. A base model (Qwen3.5-9B-Base) is also available for fine-tuning and research. Multiple quantized variants are available.

Supported inference frameworks: vLLM (recommended), SGLang (recommended), llama.cpp, MLX (Apple Silicon), Hugging Face Transformers, and KTransformers.

Strengths

• Outperforms Qwen3-30B (3x its size) on GPQA, IFEval, and long-context benchmarks
• Controls GPT-5-Nano on vision: MMMU-Pro +13, MathVision +17, OmniDocBench +32
• 262K to 1M context window from just 9B parameters - constant memory via DeltaNet
• Natively multimodal: text, images, and video without adapter overhead
• Runs on a single consumer GPU - RTX 4090 at BF16, RTX 3060 at 4-bit
• 201 languages with 248K vocabulary - strong multilingual coverage
• VideoMME 84.5 rivals models 10x its size

Weaknesses

• Trails the Qwen3.5-27B by 4-8 points on top-end benchmarks (MMLU-Pro 82.5 vs 86.1)
• All benchmarks are self-reported - independent validation pending
• Gated DeltaNet architecture requires compatible inference frameworks for ideal speed
• 9B all-active means higher per-token cost than the 35B-A3B MoE (3B active)
• No managed API specifically aligned with this model - self-hosting only
• Thinking mode adds latency and token cost for simple tasks

Related Coverage

• Qwen 3.5 Small Series Ships Four Models
• Qwen 3.5 Medium Series Drops Four Models
• Qwen3.5: 397B Parameters, 17B Active
• Qwen3.5-27B
• Qwen3.5-4B
• Open Source LLM Leaderboard

Sources:

• Qwen3.5-9B Model Card (HuggingFace)
• Qwen3.5-9B-Base Model Card (HuggingFace)
• QwenLM/Qwen3.5 GitHub Repository
• Qwen3.5: Towards Native Multimodal Agents (Blog)
• Alibaba Cloud Blog: Qwen3.5
• Gated Delta Networks Paper (arXiv)