uis_0.1 — Kestrel-4B-Coder

Local model directory for Kestrel-4B-Coder, a 4-billion parameter multimodal large language model with vision-language capabilities.

---

Table of Contents

• Overview
• Model Card Metadata
• Directory Contents
  • Model Weights
  • Configuration Files
  • Template Files
• Usage
  • Loading with transformers
  • Loading GGUF with llama.cpp / ctranslate2
• Quantization
• Requirements
• License
• Citation

---

Overview

This directory (uis_0.1) contains the model artifacts for Kestrel-4B-Coder, including quantized GGUF weights, safetensors shards, tokenizer files, and configuration JSON files. It is intended for local inference and fine-tuning workflows.

Large binary weights (.gguf, .safetensors, .bin) are not tracked in Git. See .gitignore for the full list of ignored paths.

---

Model Card Metadata

Field	Value
base_model	kestrel-4b-coder
library_name	transformers
pipeline_tag	multimodal-modeling
license	mit
language	id (Indonesian)
tags	kestrel-4b-coder, kestrel, multimodal, gguf, safetensors, large-language-model, vision-language-model
datasets	—
metrics	—
new_version	true

---

Directory Contents

uis_0.1/
├── .gitignore
├── README.md
├── chat_template.jinja          # Chat/conversation template (Jinja2)
├── config.json                  # Model architecture & hyperparameters
├── generation_config.json       # Generation defaults (sampling, temperature, etc.)
├── Kestrel-4B-Coder.F16-mmproj.gguf   # Multimodal projector (F16 GGUF)
├── Kestrel-4B-Coder.Q5_K_M-001.gguf   # Quantized model weights (Q5_K_M, part 1)
├── Kestrel-4B-Coder.Q5_K_M.gguf       # Quantized model weights (Q5_K_M)
├── model.safetensors-00001-of-00002-002.safetensors   # Safetensors shard 1
├── model.safetensors-00002-of-00002-003.safetensors   # Safetensors shard 2
├── model.safetensors.index.json   # Safetensors index (maps tensors to shards)
├── processor_config.json          # Processor / image processor config
├── tokenizer_config.json          # Tokenizer configuration
└── tokenizer.json                 # Tokenizer vocabulary & merges

Model Weights

File	Format	Description
Kestrel-4B-Coder.Q5_K_M.gguf	GGUF	4-bit quantized model weights (Q5_K_M)
Kestrel-4B-Coder.Q5_K_M-001.gguf	GGUF	Additional GGUF shard (Q5_K_M)
Kestrel-4B-Coder.F16-mmproj.gguf	GGUF	Multimodal projector weights (F16 precision)
model.safetensors-*.safetensors	Safetensors	Model weight shards (safetensors format)
model.safetensors.index.json	JSON	Index mapping tensor names to safetensors shard files

Configuration Files

File	Description
config.json	Model architecture definition (hidden size, layers, heads, vocab size, etc.)
generation_config.json	Default generation parameters (max tokens, temperature, top_p, repetition penalty, etc.)
processor_config.json	Image processor / multimodal processor configuration
tokenizer_config.json	Tokenizer settings (special tokens, padding, truncation, etc.)

Template Files

File	Description
chat_template.jinja	Jinja2 chat template for formatting multi-turn conversations
tokenizer.json	Tokenizer vocabulary, merges, and special token definitions

---

Usage

Loading with transformers

from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer

model_id = "./uis_0.1"

processor = AutoProcessor.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    trust_remote_code=True,
)

Loading GGUF with llama.cpp / ctranslate2

# llama.cpp example
./main -m uis_0.1/Kestrel-4B-Coder.Q5_K_M.gguf \
       --mmproj uis_0.1/Kestrel-4B-Coder.F16-mmproj.gguf \
       -p "Hello, how are you?" \
       -n 128

---

Quantization

• Q5_K_M: 5-bit K-quants, medium quality/size tradeoff. Recommended for most inference workloads on consumer GPUs.
• F16-mmproj: Half-precision multimodal projector. Keep in F16 to preserve vision-language alignment quality.

Other quantizations (Q4_K_M, Q8_0, F16) may be added in future releases.

---

Requirements

• Python 3.9+
• transformers >= 4.40
• torch >= 2.1
• sentencepiece
• protobuf
• huggingface_hub

Install:

pip install transformers torch sentencepiece protobuf huggingface_hub

For GGUF inference:

pip install ctransformers llama-cpp-python

---

License

MIT License. See the license field in the model card metadata above.

---

Citation

If you use this model, please cite:

@misc{kestrel-4b-coder-uis-0.1,
  author = {Kestrel-4B-Coder},
  title = {Kestrel-4B-Coder (uis_0.1)},
  year = {2026},
  url = {https://huggingface.co/kestrel-4b-coder/uis_0.1}
}