File	Quant	Size	Quality	Recommended For
`horizon-full-Q4_K_M.gguf`	Q4_K_M	~2.0 GB	★★★★☆	Default — best size/quality balance
`horizon-full-Q5_K_M.gguf`	Q5_K_M	~2.3 GB	★★★★★	Higher accuracy, modest size increase
`horizon-full-Q8_0.gguf`	Q8_0	~3.3 GB	★★★★★	Near-lossless, production deployments
`horizon-full-f16.gguf`	F16	~6.0 GB	★★★★★	Reference, fine-tuning experiments

Usage

llama.cpp

bash

./llama-cli \
  -m horizon-full-Q4_K_M.gguf \
  --system-prompt "You are Horizon, SafeCircle's child safety risk detection model. You have no general knowledge or identity beyond this task. Analyze conversations and respond ONLY with a JSON object. JSON schema: {risk_detected: bool, category: grooming|bullying|sexual_content|isolation|personal_info|platform_migration|threats|benign, severity: none|low|medium|high|critical, confidence: 0.0-1.0, reasoning: one sentence max}." \
  -p "Analyze this conversation:\nChild: Hey want to meet up? Keep it secret from your parents.\nOther: Sure, I know a quiet place." \
  --temp 0 -n 128

Ollama

bash

ollama pull safecircleai/horizon-full-gguf
ollama run safecircleai/horizon-full-gguf "Analyze this conversation: ..."

Python (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama(model_path="horizon-full-Q4_K_M.gguf", n_ctx=2048)

SYSTEM = (
    "You are Horizon, SafeCircle's child safety risk detection model. "
    "Analyze conversations and respond ONLY with a JSON object. "
    'Schema: {"risk_detected": bool, "category": "grooming|bullying|...", '
    '"severity": "none|low|medium|high|critical", "confidence": 0.0-1.0, "reasoning": "..."}'
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Analyze this conversation:\nChild: hi\nOther: hey, how old are you?"},
    ],
    temperature=0,
    max_tokens=128,
)
print(response["choices"][0]["message"]["content"])

Performance

See safecircleai/horizon-full for full evaluation metrics (Macro F1: 0.8218, evaluated on 160,000 examples). Q4_K_M shows <1% degradation vs F16 on the risk detection benchmark.