Qwen3.8-27B · CuteHuman (GGUF)

The cute-persona LoRA as GGUF — mounted at runtime by llama.cpp, never merged.

_{21 MB / f16 / 128 tensors  ·  alpha=32 embedded, runtime scale = 2.0  ·  Apache-2.0}

---

🖼 Showcase

_{Chinese · “you there?”
English · “how's it going?”
Terminal agent · restoring a deleted SVG, then writing a ramen song about it}

Same prompt, base vs CuteHuman

_{① 做一个波斯猫跳舞的 SVG 动画 — “an SVG animation of a Persian cat dancing”}

<sub>Qwen3.8-27B · generation
+ CuteHuman LoRA · generation
Qwen3.8-27B · rendered SVG
+ CuteHuman LoRA · rendered SVG</sub>

_{SVG sources: base.svg · cute.svg}

_{② make a svg animation of your favorite food}

<sub>Qwen3.8-27B · generation
+ CuteHuman LoRA · generation
Qwen3.8-27B · rendered SVG
+ CuteHuman LoRA · rendered SVG</sub>

_{SVG sources: base.svg · cute.svg}

---

📊 Before / after

_{Capabilities are essentially intact — what changes is tone and length.}

---

🚀 Usage

Prerequisite: any Qwen3.8-27B GGUF base (unsloth's Qwen3.8-27B-UD-Q4_K_M.gguf is a good pick; add mmproj-F16.gguf if you need vision).

llama-server

llama-server \
  -m Qwen3.8-27B-UD-Q4_K_M.gguf \
  --lora cutehuman-27b-lora-f16.gguf \
  --mmproj mmproj-F16.gguf \
  -c 32768 -ngl 99 -fa on --jinja \
  --alias qwen3.8-27b --host 0.0.0.0 --port 18080

llama-cli, one shot

llama-cli -m Qwen3.8-27B-UD-Q4_K_M.gguf --lora cutehuman-27b-lora-f16.gguf \
  -p "今天上班好累啊" -n 256 --temp 0.7 --top-p 0.8

Check it actually loaded

curl -s http://127.0.0.1:18080/lora-adapters
# expected: [{"id":0,"path":".../cutehuman-27b-lora-f16.gguf","scale":1.0,...}]

Toggle / rescale at runtime (no restart, no reload)

# off (equals the plain base model)
curl -X POST http://127.0.0.1:18080/lora-adapters -H 'Content-Type: application/json' -d '[{"id":0,"scale":0.0}]'
# half strength
curl -X POST http://127.0.0.1:18080/lora-adapters -H 'Content-Type: application/json' -d '[{"id":0,"scale":0.5}]'
# on
curl -X POST http://127.0.0.1:18080/lora-adapters -H 'Content-Type: application/json' -d '[{"id":0,"scale":1.0}]'

_{Or pin the scale at launch: --lora-scaled cutehuman-27b-lora-f16.gguf:0.6}

---

💡 Notes

• No system prompt was used in training (the data is user/assistant only) — leaving it empty works best.
• Turn thinking off for chit-chat: send "chat_template_kwargs": {"enable_thinking": false} in the request body.
• Sampling: temperature 0.7 / top_p 0.8 / top_k 20; for thinking mode use the official defaults 1.0 / 0.95.
• Language layers only: the LoRA never touches the vision tower, so keep loading --mmproj as usual — image understanding is unaffected.
• Do not enable MTP speculative decoding (--spec-type draft-mtp): the draft head was trained on the base model, so with the LoRA attached the hit rate collapses — measured 41.5 tok/s down to 27.8 tok/s, plus 1.3 GB of VRAM wasted.

---

📁 Companion

_{LoRA details: r=16 / alpha=32 / dropout=0, attached to q_proj, k_proj, v_proj, o_proj of the 16 full-attention layers, 10.49M trainable params. Demo images and SVG samples live in demos/. Chinese card: ModelScope}