Instructions to use nexbridgesolutions/Arden-2.0-280M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use nexbridgesolutions/Arden-2.0-280M with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nexbridgesolutions/Arden-2.0-280M:Q4_0
# Run inference directly in the terminal:
llama cli -hf nexbridgesolutions/Arden-2.0-280M:Q4_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nexbridgesolutions/Arden-2.0-280M:Q4_0
# Run inference directly in the terminal:
llama cli -hf nexbridgesolutions/Arden-2.0-280M:Q4_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf nexbridgesolutions/Arden-2.0-280M:Q4_0
# Run inference directly in the terminal:
./llama-cli -hf nexbridgesolutions/Arden-2.0-280M:Q4_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf nexbridgesolutions/Arden-2.0-280M:Q4_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf nexbridgesolutions/Arden-2.0-280M:Q4_0

Use Docker

docker model run hf.co/nexbridgesolutions/Arden-2.0-280M:Q4_0

Llama new
LM Studio
Jan

vLLM

How to use nexbridgesolutions/Arden-2.0-280M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nexbridgesolutions/Arden-2.0-280M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nexbridgesolutions/Arden-2.0-280M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nexbridgesolutions/Arden-2.0-280M:Q4_0

Ollama
How to use nexbridgesolutions/Arden-2.0-280M with Ollama:
```
ollama run hf.co/nexbridgesolutions/Arden-2.0-280M:Q4_0
```
Unsloth Desktop
Docker Model Runner
How to use nexbridgesolutions/Arden-2.0-280M with Docker Model Runner:
```
docker model run hf.co/nexbridgesolutions/Arden-2.0-280M:Q4_0
```

Lemonade

How to use nexbridgesolutions/Arden-2.0-280M with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull nexbridgesolutions/Arden-2.0-280M:Q4_0

Run and chat with the model

lemonade run user.Arden-2.0-280M-Q4_0

List all available models

lemonade list

Atomic Chat

Arden 2.0 280M — Preview 10% (step 232,503)

Arden 2.0 is a decoder-only language model trained from scratch by Nex Bridge Solutions LLC. No pretrained weights, no model forks: architecture, tokenizer, data pipeline and training loop are all built in-house in PyTorch.

Project page: ardenlm.com

This repository holds an early preview checkpoint of an ongoing training run, published in the spirit of open development.

This is a base model (pretraining only, no SFT). It continues text; it does not follow instructions or hold conversations yet. A chat template is included for compatibility, but the model has not been trained on that format.

Training progress


Checkpoint	step 232,503 of 2,325,034 (10.00%)
Tokens processed	~952M (952,332,288)
Validation loss	3.8308 (perplexity 46.1)
Validation set	25 fixed held-out batches (no shuffle), same batches across all evaluations

Loss is still decreasing steadily. Expect noticeably better text quality in later checkpoints.

Model details


Parameters	279,913,984 (~280M)
Architecture	Decoder-only Transformer (GPT-2 style), pre-LayerNorm, GELU, learned positional embeddings
Layers / Heads / Hidden	26 / 14 / 896 (head dim 64)
FFN size	3,584
Context length	512 tokens
Tied embeddings	Yes
Tokenizer	ByteLevel BPE, 32,000 tokens, NFKC normalization
Languages	English-primary, with Spanish, Portuguese and French
Training data	Multi-source pretraining corpus curated by Nex Bridge Solutions (predominantly English)
Optimizer	AdamW, lr 3e-4, cosine decay, 1,000 warmup steps
Effective batch	8 sequences × 512 tokens (4,096 tokens/step)
Training precision	float32

Files

File	Size	Use
`Arden-2.0-280M-preview-step232k-q4_0.gguf`	183 MB	Smallest, for CPU / low-memory devices
`Arden-2.0-280M-preview-step232k-q8_0.gguf`	316 MB	Recommended, near-lossless
`config.json`, `arden_meta.json`	—	Architecture and training metadata
`tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json`, `tokenizer_meta.json`	—	Tokenizer and chat template

Special tokens and chat format

Token	ID
`<PAD>`	0
`<UNK>`	1
`<BOS>`	2
`<EOS>`	3
`<MASK>`	4
`<USER>`	5
`<ASSISTANT>`	6
`<SYSTEM>`	7

Chat template (embedded in the GGUF and in tokenizer_config.json):

<BOS><SYSTEM>{system}<EOS><USER>{user}<EOS><ASSISTANT>{assistant}<EOS>

This format is reserved for the upcoming chat (SFT) release. On this base checkpoint, plain text completion gives the best results.

Usage

GPT4All

GPT4All by Nomic AI runs Arden locally on your computer, on CPU or GPU, with no internet connection required.

Download Arden-2.0-280M-preview-step232k-q8_0.gguf (or q4_0).
Place it in your GPT4All models folder and select it.
The chat template is read from the GGUF automatically.

Recommended GPT4All settings (Settings → Model):

Setting	Value
Context Length	512
Max Length	200
Prompt Batch Size	128
Temperature	0.8
Top-P	0.9
Top-K	40
Min-P	0
Repeat Penalty Tokens	256
Repeat Penalty	1.3
GPU Layers	26 (all layers)

Keep prompt + response under 512 tokens. Arden uses learned positional embeddings, so the context cannot be shifted once it is full. If the conversation exceeds 512 tokens, GPT4All may stop or close. Start a new chat when a conversation gets long.

Because this is a base model, replies will read as text continuation rather than assistant-style answers.

llama.cpp

Use -no-cnv for text completion:

./llama-completion -m Arden-2.0-280M-preview-step232k-q8_0.gguf \
  -p "The history of" -n 100 -no-cnv \
  --temp 0.8 --top-k 40 --top-p 0.9 --repeat-penalty 1.2

Important: Arden uses a custom chat template. Running llama-cli or llama-completion in their default conversation mode without --jinja will abort with this custom template is not supported. Use -no-cnv (completion) or --jinja (chat template).

Ollama

FROM ./Arden-2.0-280M-preview-step232k-q8_0.gguf
TEMPLATE """{{ .Prompt }}"""
PARAMETER stop "<EOS>"
PARAMETER num_ctx 512
PARAMETER temperature 0.8
PARAMETER top_k 40
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.2

ollama create arden2-preview -f Modelfile
ollama run arden2-preview "The history of"

Recommended sampling

Parameter	Value
temperature	0.7 – 0.8
top_k	40 – 50
top_p	0.9
repeat penalty	1.1 – 1.3

Greedy decoding (temperature 0) tends to loop at this stage of training; sampling with a repetition penalty is strongly recommended.

Limitations

Early checkpoint: about 10% of the planned training. Text is often repetitive or loses coherence over longer passages.
Base model: no instruction tuning, no safety tuning, no alignment. It may produce inaccurate, biased or inappropriate content.
Language balance: the corpus is predominantly English; Spanish, Portuguese and French output is weaker.
Short context: 512 tokens, with learned positions (no context shifting). Prompt plus response must stay within 512 tokens.
No RAG / LocalDocs support: the 512-token context cannot fit retrieved documents, and the base model has not been trained to answer from a given context. Disable LocalDocs collections in GPT4All when using Arden.
Not suitable for production use or for any decision-making.

Release plan

Release	Status
Preview — step 178,500 (7.7%)	✅ Released (tag `preview-step-178k`)
10% — step 232,503	✅ This release
50% — step 1,162,517	📋 Planned
100% — step 2,325,034	📋 Planned
Arden 2.0 Chat (SFT)	📋 Planned

Earlier release: Arden 1.0 280M (tag v1.0).

Custom training & fine-tuning

The GGUF files in this repository show how Arden is trained and how it performs. For business use, we offer:

Fine-tuning on your own data, starting from the trained Arden checkpoints
Domain adaptation for your industry or language needs
Assistant / persona training (SFT) for chat applications
Private deployment: on-premise or private API, with integration and support

Full-precision weights are available to clients under a commercial agreement.

Contact: sales@nexbridgesolution.com · ardenlm.com

License

Released under the Arden Community License v1.0.

The GGUF files in this repository may be used for:

✅ Personal use, research and education
✅ Internal evaluation

A commercial agreement with Nex Bridge Solutions LLC is required to:

❌ Host Arden as a service
❌ Embed Arden in commercial products
❌ Create fine-tuned or derivative models for commercial use

Commercial licensing: legal@nexbridgesolution.com

Citation

@misc{arden2026,
  title  = {Arden 2.0: A Language Model Trained From Scratch},
  author = {{Nex Bridge Solutions LLC}},
  year   = {2026},
  publisher = {Nex Bridge Solutions LLC},
  url    = {https://huggingface.co/nexbridgesolutions/Arden-2.0-280M}
}

Downloads last month: 894

GGUF

Model size

0.3B params

Architecture

gpt2

Hardware compatibility

4-bit

8-bit