Qwen3.5 and Qwen3.8-Flash-Next model cards and nested text configs (huggingface.co/Qwen)

Qwen3.5-35B-A3B card states 35B total / 3B activated, 8 routed + 1 shared expert. Both Qwen3.5 and Qwen3.8 moved to a nested multimodal config (text_config + vision_config), and both moved to a hybrid linear-attention stack: layer_types marks 3 linear_attention per 1 full_attention (30+10 for Qwen3.5-35B-A3B, 36+12 for Qwen3.8-Flash-Next), linear_conv_kernel_dim 4, full_attention_interval 4. Qwen3.8-Flash-Next card states '125B with 6B activated, plus 51B n-gram embedding and 4B MTP', which explains the 180B safetensors total. Qwen3.8-Flash-Next is under Qwen Community License 1.0, while Qwen3.5 is Apache-2.0.