Qwen/Qwen3.5-9B

Qwen3.5-9B is positioned as a dense multimodal reasoning model within the broader Qwen3.5 family, bringing together text, image, and video understanding in a single 9 billion parameter package. According to the Together AI model card, the architecture combines a hybrid Gated DeltaNet and Gated Attention design intended for efficient inference, and the model keeps a native context window that can be extended well beyond its base length with RoPE scaling. Vision is handled by a dedicated encoder that feeds into the same backbone, and the description explicitly highlights early fusion training on multimodal tokens, large-scale reinforcement learning across million-agent environments, and multi-token prediction training as part of the recipe.

In practical terms, the model is aimed at agentic and long-context workloads rather than narrow single-task use. Together AI reports function-calling results of 66.1% on BFCL-V4 and 79.1% on TAU2-Bench for production agents, alongside multimodal reasoning scores such as 89.2% on OCRBench, 84.5% on VideoMME, and 78.9% on MathVision, plus broad language coverage at 81.2% on MMMLU. The Together card also highlights an explicit thinking mode that produces reasoning traces before final answers. Because it is open weights and small enough to run locally with modest memory, as reflected in the LM Studio listing, it is a reasonable fit for teams that want a self-hostable model capable of grounded reasoning over long documents, tool-driven workflows, and multimodal inputs without the cost of a flagship-scale system.