Full merged checkpoint (base + LoRA fused, bf16 safetensors, ~8.5 GB) of kurisu-lora-v1.1-qwen3.5-4b. Use this when you want a single deployable model with no PEFT dependency — e.g. GGUF conversion for llama.cpp/Ollama, vLLM without LoRA support, or 8-bit serving on consumer GPUs.


Base	Qwen/Qwen3.5-4B
Precision	bf16 (quantize to taste)
Eval loss	1.756 (adapter's best checkpoint)

Serving on an 8 GB GPU (RTX 2080-class)

Recommended: download kurisu-4b-Q6_K.gguf from this repo and serve it with a current llama.cpp build. Q6_K is ~3.5 GB and preserved substantially more persona voice and reply depth than bitsandbytes int8 in live A/B tests:

bash

llama-server -m kurisu-4b-Q6_K.gguf \
  --ctx-size 12288 -ngl 999 \
  --repeat-penalty 1.15 --repeat-last-n 512 \
  --alias kurisu

Alternative: 8-bit with transformers + bitsandbytes (~4.5 GB weights):

python

import transformers
from transformers import AutoTokenizer, AutoConfig, BitsAndBytesConfig

MODEL = "s-g-labs/kurisu-v1.1-qwen3.5-4b-merged"
cfg = AutoConfig.from_pretrained(MODEL, trust_remote_code=True)
cls = getattr(transformers, cfg.architectures[0])
tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = cls.from_pretrained(
    MODEL, device_map="auto", trust_remote_code=True,
    quantization_config=BitsAndBytesConfig(load_in_8bit=True))

Q4_K_M (~2.6 GB) may fit larger KV caches but loses more quality. Q6_K fits comfortably on an 8 GB RTX 2080 SUPER at 12k context.

Prompting contract (ChatML, [HH:MM] bubbles, [[sticker:ref]] markers) matches the 27B card — see kurisu-lora-v1.1-qwen3.6-27b.