Qwen3.5-9B

Qwen3.5-9B is a 9-billion-parameter dense model positioned as a unified vision-language foundation, designed to bring text, image, and video understanding into a single training pipeline. Its architecture pairs gated DeltaNet layers with gated attention to balance throughput and latency, avoiding the heavier sparse mixture-of-experts pattern while still aiming for high efficiency. Early fusion across modalities lets the model reach parity with the previous Qwen3 generation and outperform the dedicated Qwen3-VL line on reasoning, coding, agent, and visual understanding benchmarks. With 201 languages of coverage and a 262K-token native window that can extend beyond a million tokens via RoPE scaling, the design intent is a compact but globally capable model that handles long documents, code, and visual inputs at inference speeds suited to production and local use alike.

The model is delivered as a post-trained checkpoint whose training methodology emphasizes early multimodal fusion, multi-token prediction pretraining, and reinforcement learning scaled across million-agent environments to build robust tool use and real-world adaptability. Native function calling is a headline strength, with strong BFCL-V4 and TAU2-Bench numbers alongside multimodal scores such as high OCRBench, VideoMME, and MathVision results, plus a thinking mode that produces explicit reasoning traces for harder problems. Open weights, compatibility with Hugging Face Transformers, vLLM, SGLang, and KTransformers, and support for LoRA fine-tuning make it a flexible base for custom deployments. In practice it fits well as a multimodal reasoning and agent backbone for production APIs, and as a sweet-spot local model for developers who want a quality step up from smaller chat models without committing to the cost of much larger frontier systems.