Qwen3.8 27B

27.8B parameter open-weight model. The dense flagship of the Qwen3.8 generation — 27.8B parameters, 256K context, image input, Apache 2.0. Architecturally a hybrid rather than a plain GQA stack: config.json declares a full-attention layer every fourth block with linear-attention and gated-delta components between them, so ordinary KV-cache arithmetic does not transfer and no architectureNotes are published for this record. The Ollama library Q4_K_M is 18 GB, Q8_0 is 30 GB and BF16 is 56 GB; with the 1.5 GB runtime overhead this site grades against, Q4 asks 19.5 GB, so 20 GB is a tight fit and 24 GB clears it comfortably. Shares a parameter count and architecture string with Qwen3.6 27B and is a different checkpoint — do not read one record's measurements onto the other. No throughput measured on our rigs yet. An official FP8 checkpoint (Qwen/Qwen3.8-27B-FP8) exists for vLLM and SGLang serving, which is a server path, not a llama.cpp one. Figures verified against the Hugging Face model card, its config.json, and the Ollama library on 2026-08-19.