Field	Value
Source checkpoint	`ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth`
Base model	`kai-os/Carnice-9b`
Dataset	`ermiaazarkhalili/Claude-Opus-Reasoning`
Training	N=1 full epoch (7,500 steps, effective batch=8)
Conversion	Unsloth `save_pretrained_gguf` → llama.cpp GGUF
Quantization tool	llama.cpp `llama-quantize`

Available quantizations

File	Size	Notes
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q2_K.gguf`	smallest	2-bit; extreme compression, quality loss
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q3_K_M.gguf`	small	3-bit; modest quality trade-off
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q4_K_M.gguf`	recommended	4-bit; best size/quality balance
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q5_K_M.gguf`	balanced	5-bit; near-full quality
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q6_K.gguf`	high quality	6-bit; minimal degradation
`Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth.Q8_0.gguf`	largest	8-bit; closest to bf16 source

Recommended default: Q4_K_M (4-bit, K-quant medium). For memory-constrained deployment, try Q2_K or Q3_K_M. For maximum fidelity, use Q8_0.

Usage

llama.cpp

bash

# Text-only
llama-cli -hf ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF --jinja -p "Explain step-by-step how to compute 23 × 47 without a calculator." -n 256

# Interactive chat
llama-cli -hf ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF --jinja -cnv

Ollama

ollama run hf.co/ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF:Q4_K_M

llama-cpp-python

python

from llama_cpp import Llama
llm = Llama.from_pretrained(
    repo_id="ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF",
    filename="*Q4_K_M.gguf",
    n_ctx=2048,
)
out = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Explain step-by-step how to compute 23 × 47 without a calculator."}],
    max_tokens=256,
)
print(out["choices"][0]["message"]["content"])

Intended use

For research and non-commercial experimentation only. Outputs should be independently verified before any downstream use.

Limitations

GGUF quantizations have unavoidable quality loss relative to the source bfloat16 checkpoint. Use Q5_K_M or Q8_0 for best fidelity.
Inherits all limitations of the source merged checkpoint (ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth).
Reasoning distilled from Claude Opus; quality bounded by teacher trace fidelity on the 60k-example distillation set.

Citation

bibtex

@misc{ carnice_9b_sft_reasoning_unsloth_2026_gguf ,
  author = {Ermia Azarkhalili},
  title = { Carnice-9B-SFT-Reasoning-Unsloth — GGUF quantized },
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ermiaazarkhalili/Carnice-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF}}
}

This gemma2 model was trained 2× faster with Unsloth and Hugging Face's TRL library.