# 16 GiB VRAM (primary desktop GPU, single user) # /home/hans/ai/models/llm_presets/16gb.ini [*] flash-attn = on gpu-layers = all parallel = 1 cache-type-k = q4_0 cache-type-v = q4_0 batch-size = 384 ubatch-size = 384 load-mode = mlock spec-type = draft-mtp,ngram-mod spec-draft-n-max = 2 spec-ngram-mod-n-match = 24 spec-ngram-mod-n-min = 24 spec-ngram-mod-n-max = 32 image-min-tokens = 1024 image-max-tokens = 2048 # KV cache VRAM # Qwen3.8-27B: 16 attn layers × 4 KV heads × 256 dim × 2 (K+V) # q4_0/q4_0: 0.5 B/elem → 16 KiB/token → GiB ≈ ctx / 65536 # https://datakeskus.io/files/qwen38-27b-kv-cache.txt #ctx-size = 2048 # 0.03 GiB #ctx-size = 4096 # 0.06 GiB #ctx-size = 6144 # 0.09 GiB #ctx-size = 8192 # 0.13 GiB #ctx-size = 16384 # 0.25 GiB #ctx-size = 32768 # 0.50 GiB #ctx-size = 49152 # 0.75 GiB #ctx-size = 65536 # 1.00 GiB #ctx-size = 81920 # 1.25 GiB #ctx-size = 98304 # 1.50 GiB #ctx-size = 114688 # 1.75 GiB #ctx-size = 131072 # 2.00 GiB #ctx-size = 147456 # 2.25 GiB #ctx-size = 163840 # 2.50 GiB #ctx-size = 180224 # 2.75 GiB #ctx-size = 196608 # 3.00 GiB #ctx-size = 229376 # 3.50 GiB #ctx-size = 245760 # 3.75 GiB #ctx-size = 262144 # 4.00 GiB # Model VRAM # Qwen3.8-27B-ZB4.00-MIN-v3-IQ4_XS.gguf 12.7 GiB # mmproj-Qwen3.8-27B-BF16.gguf 0.89 GiB # https://huggingface.co/tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF # https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF [Instruct] ctx-size = 32768 temperature = 0.7 top-p = 0.8 top-k = 20 min-p = 0.0 presence-penalty = 1.5 repeat-penalty = 1.0 reasoning = off chat-template-kwargs = {"enable_thinking": false} [Instruct Vision CPU] ctx-size = 32768 mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-BF16.gguf # saves VRAM; slower image/PDF/video analysis no-mmproj-offload = on temperature = 0.7 top-p = 0.8 top-k = 20 min-p = 0.0 presence-penalty = 1.5 repeat-penalty = 1.0 reasoning = off chat-template-kwargs = {"enable_thinking": false} [Instruct Vision GPU] ctx-size = 32768 mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-BF16.gguf temperature = 0.7 top-p = 0.8 top-k = 20 min-p = 0.0 presence-penalty = 1.5 repeat-penalty = 1.0 reasoning = off chat-template-kwargs = {"enable_thinking": false} [Thinking] ctx-size = 32768 m = /home/hans/ai/models/llm/Qwen3.8-27B-ZB4.00-MIN-v3-IQ4_XS.gguf temperature = 1.0 top-p = 0.95 top-k = 20 min-p = 0.0 presence-penalty = 0.0 repeat-penalty = 1.0 reasoning-preserve = on reasoning-effort = medium chat-template-kwargs = {"reasoning_effort": "medium"} [Thinking Vision CPU] ctx-size = 32768 mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-BF16.gguf # saves VRAM; slower image/PDF/video analysis no-mmproj-offload = on temperature = 1.0 top-p = 0.95 top-k = 20 min-p = 0.0 presence-penalty = 0.0 repeat-penalty = 1.0 reasoning-preserve = on reasoning-effort = medium chat-template-kwargs = {"reasoning_effort": "medium"} [Thinking Vision GPU] ctx-size = 32768 mmproj = /home/hans/ai/models/llm_aux/mmproj-Qwen3.8-27B-BF16.gguf temperature = 1.0 top-p = 0.95 top-k = 20 min-p = 0.0 presence-penalty = 0.0 repeat-penalty = 1.0 reasoning-preserve = on reasoning-effort = medium chat-template-kwargs = {"reasoning_effort": "medium"}