Qwen3.8 27B

27B dense vision-language model by Alibaba, Apache-2.0, reading images and video alongside text. Its layer stack is hybrid — 48 of the 64 layers run Gated DeltaNet linear attention and only 16 keep a KV cache — so context costs roughly a quarter of what the dimensions suggest: 32K works out at about 2.1 GB where an all-attention 27B of the same shape would spend 8.6 GB, and a 128K window costs about what that model would spend on 32K. The full 262,144-token context still wants around 17 GB of KV on top of the weights. Ollama serves it as qwen3.8:27b at 18 GB with vision, tools and thinking enabled; llama.cpp runs ggml-org's own GGUF alongside a 0.93 GB mmproj projector for the image path. The 4-bit footprint depends on who built it — Q4_K_M spans 16.81 GB (lmstudio-community) to 18.97 GB (ggml-org).