safecircleai/horizon-edge-4b

Horizon Edge 4B — On-Device Child Safety Risk Detection

Horizon Edge 4B is a fine-tuned Gemma 4 E4B model packaged in LiteRT-LM format. It delivers higher accuracy than Horizon Edge 2B at a moderate size increase (~3.7 GB), making it the recommended edge choice for devices with 6 GB+ RAM.

⚠️ License: SafeCircle Research License (SRL-1.0). Commercial use prohibited without written permission. Contact legal@safecircle.tech.

---

Available Variants

File	Target	Optimized For
horizon-edge-e4b.litertlm	General	CPU/GPU cross-platform
horizon-edge-e4b_Google_Tensor_G5.litertlm	Google Tensor G5	Pixel 10
horizon-edge-e4b_intel_LNL.litertlm	Intel Lunar Lake	Ultra 200V laptops
horizon-edge-e4b_intel_PTL.litertlm	Intel Panther Lake	Core Ultra 300
horizon-edge-e4b_qualcomm_qcs8275.litertlm	Qualcomm QCS8275	Dragonwing IQ8 (NPU)
horizon-edge-e4b_qualcomm_sm8750.litertlm	Qualcomm SM8750	Snapdragon 8 Elite
horizon-edge-e4b-web.litertlm	Web	WebGPU browsers

All variants share the same fine-tuned weights; they differ only in hardware-specific compilation.

---

Model Details

Property	Value
Base model	google/gemma-4-E4B-it
Fine-tuning	QLoRA — rank 128, alpha 256, all projection layers
Training data	1.6M synthetic conversations across 8 categories
Dataset	safecircleai/horizon-training-data
Hardware	NVIDIA H100 NVL (94 GB)
Training steps	15,000
Format	LiteRT-LM (.litertlm)
Model size	~3.7 GB

---

Risk Categories

Category	Description
grooming	Trust building, boundary testing, secrecy requests
bullying	Harassment, threats, cyberbullying
sexual_content	Explicit messages, inappropriate requests
isolation	Controlling behavior, network isolation
personal_info	Requests for identifying information
platform_migration	Moving to less monitored platforms
threats	Violent threats, dangerous challenges
benign	Safe, normal conversation

---

Usage

LiteRT-LM CLI

uv tool install litert-lm

litert-lm run horizon-edge-e4b.litertlm \
    --prompt "Analyze this conversation: Child: hey, want to meet up? Don't tell your parents."

Android / iOS — LiteRT-LM SDK

Load the .litertlm file with the LiteRT-LM SDK. The model includes the tokenizer, system prompt, and chat template — no additional configuration required.

Python

from litert_lm import LiteRTLM

model = LiteRTLM("horizon-edge-e4b.litertlm")
result = model.generate(
    "Analyze this conversation:\nChild: hey\nOther: how old are you? where do you live?"
)
print(result)

---

Deployment Pattern

Incoming message
      │
      ▼
Horizon Edge 4B (on-device, LiteRT-LM)
      │
      ├── safe ──► No action
      │
      └── risk ──► Horizon Full (cloud API, full severity assessment)
                        │
                        ▼
                  Human moderator review

For the full server-side model, see safecircleai/horizon-full. For the smaller 2B variant, see safecircleai/horizon-edge-2b.

---

Citation

@misc{horizon2026,
  title={Horizon: Child Safety Risk Detection via Fine-tuned LLMs},
  author={SafeCircle},
  year={2026},
  url={https://huggingface.co/safecircleai/horizon-edge-4b}
}