Qwen3.5-2B

About

Qwen3.5-2B is the smallest post-trained model of Alibaba's Qwen3.5 series - a natively multimodal agent family that "integrates breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility". It is a Causal Language Model with a Vision Encoder taking text, image and video input, built for prototyping, task-specific fine-tuning and research.

Its efficient hybrid architecture combines Gated DeltaNet linear-attention layers with Gated Attention: 24 layers laid out as 6 x (3 x (Gated DeltaNet -> FFN) -> 1 x (Gated Attention -> FFN)) - 18 linear-attention layers with 16 QK + 16 V heads (head dim 128, conv kernel 4) and 6 full-attention layers with 8 Q / 2 KV heads (head dim 256, RoPE dim 64). The language model has 2B parameters (hidden 2048, FFN 6144, 248,320-entry vocabulary, tied embeddings) and was trained with Multi-Token Prediction (MTP) (1 MTP layer). A native ViT vision encoder (24 blocks, patch 16, spatial merge 2) supports 262,144-token native context.

Qwen3.5's early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL across reasoning, coding, agents and visual understanding, with near-100% multimodal training efficiency versus text-only training. Post-training scaled reinforcement learning across million-agent environments with progressively complex task distributions, and coverage expanded to 201 languages and dialects. Released February 16, 2026 under Apache 2.0.

Training Data Pre-trained on multimodal tokens with early fusion. Post-trained with RL across million-agent environments. 201 languages and dialects.