lapa-v0.1.3-instruct-Q8_0-GGUF

Lapa v0.1.3 Instruct — Q8_0 GGUF

8-bit (Q8_0) GGUF quantization of lapa-llm/lapa-v0.1.3-instruct — a Ukrainian-specialized instruction model built on Google's Gemma 3 12B, with roughly 80 000 of its 256 000 tokenizer tokens replaced by Ukrainian-oriented ones.

This repository contains only the quantized weights. For the original model, tokenizer sources, training details and intended use, see the upstream model card.

Files

Why Q8_0

Q8_0 is the highest-fidelity common GGUF quant (8-bit, ~0.5 GB per B params). It keeps generation quality effectively indistinguishable from the bf16 original while roughly halving memory, so the full 12B model fits comfortably on one 24 GB card (e.g. RTX 3090/4090) with room for KV cache. This matters here because Lapa's modified tokenizer makes the model sensitive to precision loss — lower-bit quants (Q4/Q5/Q6) risk degrading Ukrainian morphology, so Q8_0 is the recommended operating point.

How it was made

Following the official Lapa quantization recipe, which uses a patched converter that preserves Lapa's custom tokenizer (tokenizer.model + merge_info.json) so no vocabulary precision is lost:

# llama.cpp @ b8185
# 1. HF safetensors -> F16 GGUF (patched Lapa converter)
python convert_hf_to_gguf_lapa.py lapa-v0.1.3-instruct/ \
    --outfile lapa-12B-it-F16.gguf --outtype f16

# 2. F16 -> Q8_0
llama-quantize lapa-12B-it-F16.gguf lapa-v0.1.3-instruct-Q8_0.gguf Q8_0

Usage

llama-server (OpenAI-compatible API)

llama-server \
  --model lapa-v0.1.3-instruct-Q8_0.gguf \
  --host 0.0.0.0 --port 8080 \
  --n-gpu-layers 999 --ctx-size 8192 --flash-attn on

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Розкажи коротко про місто Львів."}],
       "temperature":0.3}'

llama-cli

llama-cli -m lapa-v0.1.3-instruct-Q8_0.gguf -ngl 999 -c 8192 \
  -p "Перепиши речення без цифр: На 500 сторінках є примітки."

Download

huggingface-cli download skypro1111/lapa-v0.1.3-instruct-Q8_0-GGUF \
  lapa-v0.1.3-instruct-Q8_0.gguf --local-dir .

License

Derived from Google Gemma 3 and inherits the Gemma Terms of Use. Use of these weights is subject to those terms and to any terms of the upstream lapa-llm/lapa-v0.1.3-instruct model.

Credits

• Original model: Lapa LLM team — lapa-llm/lapa-v0.1.3-instruct
• Base architecture: Google Gemma 3
• Quantization tooling: llama.cpp

This is an unofficial, community quantization. Not affiliated with the Lapa LLM team or Google.