Kona2-small-3.8B is a compact 3.8-billion parameter Georgian language model built on Microsoft Phi-3.5-mini-instruct. It goes through the full training pipeline (pre-training + SFT + DPO) like the 12B models, but provides Georgian language capabilities with significantly lower compute requirements.

Model Summary

Property	Value
Parameters	3.8B
Architecture	Phi-3 (Transformer)
Context Length	8K tokens
Languages	Georgian (ka), English (en), other (limited)
Training	Full pipeline (Pre-training + SFT + DPO)
Vocabulary	Extended (~20K Georgian tokens)
Base Model	microsoft/Phi-3.5-mini-instruct

Intended Uses

Primary Use Cases

Edge deployment and mobile applications
Low-latency conversational AI
Georgian text generation on consumer hardware
Translation (especially strong)
Educational and research purposes
Rapid prototyping and development

Training

Training Pipeline

Same full pipeline as the 12B models, applied to the smaller Phi-3.5 base:

Vocabulary Expansion: Added ~20K Georgian tokens (1.9 tokens/word fertility)
Continue Pre-training: LoRA/DoRA on Georgian/English corpus
SFT (Supervised Fine-Tuning): Instruction tuning on Georgian instructions
DPO (Direct Preference Optimization): Preference alignment for better responses

Training Configuration

Base Model: microsoft/Phi-3.5-mini-instruct
Method: LoRA with DoRA enabled
Pre-training Context: 8K tokens
New Tokens: ~20K Georgian tokens
Precision: BF16
Infrastructure: NVIDIA H100 GPUs

Usage

Installation

pip install transformers torch accelerate

Basic Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "tbilisi-ai-lab/kona2-small-3.8B",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True  # Required for Phi-3 architecture
)
tokenizer = AutoTokenizer.from_pretrained(
    "tbilisi-ai-lab/kona2-small-3.8B",
    trust_remote_code=True
)

messages = [
    {"role": "user", "content": "გამარჯობა! რა არის ხელოვნური ინტელექტი?"}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    inputs, 
    max_new_tokens=256, 
    temperature=0.7,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Ollama (Local Deployment)

bash

# If using GGUF quantized version
ollama run kona2-small

>>> გამარჯობა!
გამარჯობა! როგორ შემიძლია დაგეხმარო?

Comparison with 12B Models

Feature	Kona2-small-3.8B	Kona2-12B
Parameters	3.8B	12B
VRAM (FP16)	~8GB	~24GB
VRAM (4-bit)	~3GB	~8GB
Speed	Faster	Slower
Quality	Good	Better
Function Calling	Basic	Full
Reasoning	Limited	Strong

When to Use Kona2-small

Running on consumer GPUs (RTX 3060, RTX 4070, etc.)
Mobile or edge deployment
High-throughput, low-latency requirements
Simple Q&A and chat applications
Development and prototyping

When to Use Kona2-12B

Production applications requiring high quality
Complex reasoning tasks
Reliable function calling
When compute resources are available

Related Models

Model	Parameters	Use Case
kona2-12B	12B	Production (DPO-aligned)
kona2-12B-Instruct	12B	Production (SFT)
kona2-12B-Base	12B	Fine-tuning

Limitations

Training data cutoff: 2024

Technical Specifications

Precision: BF16/FP16 supported
Minimum VRAM: 8GB (FP16), 3GB (4-bit)
Custom Code: Required (trust_remote_code=True)

Citation

bibtex

@misc{tbilisi2025kona2small,
  title        = {Kona2-small-3.8B: A Compact Georgian Language Model},
  author       = {Tbilisi AI Lab Team},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/tbilisi-ai-lab/kona2-small-3.8B}}
}

License

This model is released under the Apache 2.0 License.

Contact

Organization: Tbilisi AI Lab
Website: ailab.ge
Chat: chat.ailab.ge
API: api.ailab.ge