output = llm(
"Once upon a time,",
max_tokens=512,
echo=True
)
print(output)ARDEN AI
A multilingual language model built from scratch — ES / EN / PT / FR
The Arden project aims to pretrain a family of multilingual language models (Spanish, English, Portuguese, French) built entirely from scratch in PyTorch — no pretrained weights, no model forks. The current model, ARDEN 1.0 (280M parameters), is being pretrained on consumer hardware: a single NVIDIA GTX 1660 Super with 6GB of VRAM.
Yes, you read that right — 6GB. Arden is proof that you don't need a datacenter to start building language models.
Arden is developed and owned by Nex Bridge Solutions LLC (Arizona, USA), led by David Ernesto Arriaga Pineda.
How to use
With Ollama
# Download and run directly
ollama run devdavwork/Arden-1.0-280M
With llama.cpp
./llama-completion -m arden-280m-f16.gguf \
-p "Once upon a time, in a small village," \
-n 100 --temp 0.7
With Python (llama-cpp-python)
from llama_cpp import Llama
llm = Llama(model_path="arden-280m-f16.gguf", n_ctx=512)
output = llm("Once upon a time, in a small village,", max_tokens=100)
print(output["choices"][0]["text"])
News
- 2026-07: Arden 1.0 (280M) — First public GGUF release. ByteLevel BPE tokenizer, 57k training steps on 23GB proprietary corpus. Available on Hugging Face and Ollama.
- 2026-06: Migrated training from CPU to GPU (GTX 1660 Super). Full corpus processed — 23GB across 50 shards.
- 2026-06: Initial release of the Arden architecture, tokenizer, data pipeline and training loop. Licensed under the Arden Community License v1.0.
Releases Schedule
Base models:
| Phase | Checkpoint | Params | Hardware | Status |
|---|---|---|---|---|
| 1 | Arden-1.0-280M | 280M | GTX 1660 Super 6GB | ✅ Released |
| 2 | Arden-1.0-500M | ~500M | GTX 1660 Super 6GB | 📋 Planned |
| 3 | Arden-1.1B | ~1.1B | 12GB+ VRAM GPU | 📋 Planned |
Specialized models:
| Checkpoint | Params | Purpose | Status |
|---|---|---|---|
| ArdenWard | 22M | Cybersecurity event selector | 📋 Planned |
Training Details
| Setting | Description |
|---|---|
| Parameters | 308,585,984 (~280M) |
| Architecture | Decoder-only Transformer (GPT-style), pre-LayerNorm, GELU |
| Layers / Heads | 26 layers, 14 heads, d_model=896, d_ff=3584 |
| Sequence Length | 512 |
| Tokenizer | ByteLevel BPE, 32,000 vocab, trained on multilingual corpus |
| Training Steps | 57,000 (ongoing) |
| Training Data | arden_corpus_v1 (proprietary corpus, ~23GB, 50 shards, curated by Nex Bridge Solutions LLC) |
| Precision | float32 training, F16 GGUF export |
| Hardware | 1× NVIDIA GTX 1660 Super 6GB |
Project Structure
arden/
├── core/
│ ├── attention.py # Multi-head self-attention
│ ├── config.py # Master configuration
│ ├── model.py # ArdenModel — decoder-only Transformer
│ └── tokenizer.py # ByteLevel BPE tokenizer
├── data/
│ └── preprocessor.py # Tokenization & train/val/test splits
├── tools/
│ └── arden_to_gguf.py # PyTorch → GGUF converter
├── train.py # Training loop
└── LICENSE # Arden Community License v1.0
Potential Use Cases
- Local, private text generation in Spanish and English — no internet connection required.
- Deployment on modest hardware — runs entirely on CPU via llama.cpp.
- A reference codebase for anyone who wants to pretrain a language model from scratch in pure PyTorch.
- A foundation for Latin American AI — Arden treats Spanish as a first-class language.
Limitations
This is an early-stage base language model trained on 57k steps. It:
- Does not follow instructions (it is a text completion model, not a chat model)
- May produce incoherent or repetitive text
- Is not aligned or fine-tuned for safety
- Is intended for research and experimentation
License
Arden is released under the Arden Community License v1.0:
- ✅ Free for personal use, research, and education
- ✅ Free for internal evaluation
- ❌ Hosting Arden as a service requires a commercial agreement
- ❌ Embedding Arden in commercial products requires a commercial agreement
For commercial licensing: legal@nexbridgesolutions.com
About
Arden is created and owned by Nex Bridge Solutions LLC (Arizona, USA), led by David Ernesto Arriaga Pineda.
Citation
@misc{arriaga2026arden,
title = {Arden: A Multilingual Language Model Built From Scratch on Consumer Hardware},
author = {Arriaga Pineda, David Ernesto},
year = {2026},
url = {https://github.com/nexbridgesolutions/Arden}
}
Copyright 2026 Nex Bridge Solutions LLC
"Building AI from scratch, one token at a time."
- Downloads last month
- 2
16-bit
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="nexbridgesolutions/Arden-1.0-280M", filename="arden-280m-f16.gguf", )