Instructions to use automatedigital/Paradigm-2.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use automatedigital/Paradigm-2.0 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 automatedigital/Paradigm-2.0:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf automatedigital/Paradigm-2.0:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf automatedigital/Paradigm-2.0:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf automatedigital/Paradigm-2.0:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf automatedigital/Paradigm-2.0:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf automatedigital/Paradigm-2.0:Q4_K_M

Use Docker

docker model run hf.co/automatedigital/Paradigm-2.0:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use automatedigital/Paradigm-2.0 with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/automatedigital/Paradigm-2.0:Q4_K_M

Ollama
How to use automatedigital/Paradigm-2.0 with Ollama:
```
ollama run hf.co/automatedigital/Paradigm-2.0:Q4_K_M
```
Unsloth Desktop

How to use automatedigital/Paradigm-2.0 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf automatedigital/Paradigm-2.0:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "automatedigital/Paradigm-2.0:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use automatedigital/Paradigm-2.0 with Docker Model Runner:
```
docker model run hf.co/automatedigital/Paradigm-2.0:Q4_K_M
```

Lemonade

How to use automatedigital/Paradigm-2.0 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull automatedigital/Paradigm-2.0:Q4_K_M

Run and chat with the model

lemonade run user.Paradigm-2.0-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use automatedigital/Paradigm-2.0 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf automatedigital/Paradigm-2.0:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default automatedigital/Paradigm-2.0:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use automatedigital/Paradigm-2.0 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf automatedigital/Paradigm-2.0:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "automatedigital/Paradigm-2.0:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Paradigm 2.0

A standalone 27B coding model from Automate Digital, based on Swift 1.5 and Qwen3.8.

Download Paradigm-2.0-Q4_K_M.gguf and run it directly with a compatible llama.cpp server. No separate LoRA or base-model download is required. This is an experimental merged fine-tuned release.

Credits and lineage

Qwen3.8-27B — Alibaba Cloud / Qwen: underlying model, architecture and tokenizer. Copyright 2026 Alibaba Cloud; Apache 2.0. Upstream.
Swift 1.5 — UkisAI: the immediate fine-tuned base, Swift-1.5-Qwen3.8-27b, revision b4c84d42903a8646b25857eb2827288d92ed85a4. Swift Open License v1.0.
Automate Digital: Engine agentic fine-tuning, standalone merge, validation and Paradigm 2.0 packaging. Original adapter.
NVIDIA Open-SWE-Traces contributors: training data, CC BY 4.0, with underlying source attributions retained in DATA_ATTRIBUTION.json.

The name Paradigm 2.0 identifies our derivative release; it does not imply authorship of the upstream architecture or endorsement by Qwen or UkisAI.

What was trained and merged

The original training used QLoRA, rank 8 / alpha 8, for 128 optimizer updates at maximum sequence length 4096. It selected update 128. The filtered corpus contains 750 training and 85 development action windows, split by repository; these are not 750 complete agent trajectories. Training took approximately 20.3 minutes on one RTX 3090; the complete maintenance window was approximately 26 minutes.

This distribution contains complete inference weights with the learned adapter incorporated. It was not trained using full-parameter fine-tuning, and it is not a BF16-source merge or Transformers safetensors checkpoint.

The build merged the F16 GGUF adapter into the existing Swift Q4_K_M GGUF, then requantized the adapted tensors to their original tensor types. Untouched tensors, including the native MTP head, were byte-verified against the input quant. Requantization may alter or erase small learned updates. The canonical PEFT adapter remains available separately for higher-precision reconstruction.

Input base quant SHA256: 2ebba0ff1e63c1ac3fadd4e83efcea189f47f33ec72c91877af94de6ebe30590. Input adapter GGUF SHA256: c9df0cab4e375331c800a22ad01a0adea7f1d5c646b35076ebf0b7234edbce1a. The quant publisher declares Swift lineage; an exact conversion-input manifest for the original publisher quant is unavailable. See the included provenance.

Validation and limitations

The merged model passed native loading, automatic and forced tool-call checks, and the completed four-block native decode pilot on Engine GPU0. Tensor names, shapes, types and runtime metadata were checked. Branding changed only model name and description, with tensor payloads verified unchanged.

The standalone merged model has not completed the full agentic coding suite. Do not transfer base-plus-adapter coding scores to this requantized merged artifact or claim a general capability improvement. The separate adapter's matched coding evaluation and this model's native runtime evidence are documented separately in evaluation/.

This small fine-tuning run may regress reasoning, tool use or general capability. Native speculative decoding was exercised in the pilot; broader correctness and performance remain to be measured. This release is intended for local evaluation, with independent tests and review of generated changes.

Download and run

hf download automatedigital/Paradigm-2.0 Paradigm-2.0-Q4_K_M.gguf --local-dir ./Paradigm-2.0
llama-server   --model ./Paradigm-2.0/Paradigm-2.0-Q4_K_M.gguf   --alias paradigm-2.0   --host 127.0.0.1 --port 8000   --ctx-size 131072 --parallel 1 --n-gpu-layers 999   --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0   --batch-size 2048 --ubatch-size 512   --temp 1.0 --top-p 0.95 --top-k 20 --jinja

Use a llama.cpp build supporting this model's qwen35 architecture. The Engine build tested here is commit 50569eb87df530daff11afda229ceb9ab8e6cae8. For that build's optional native MTP use --spec-type draft-mtp --spec-draft-n-max 2; flag names can differ in other versions. Do not add --lora to this standalone model.

For Pi/OpenCode use an OpenAI-compatible endpoint ending /v1, model ID paradigm-2.0, and the server's configured API key. Bind publicly only with appropriate network controls and authentication. Match the existing Engine settings when comparing; changing context or sampling changes the experiment.

A single RTX 3090 was used for the native runtime tests. Available VRAM and runtime support determine usable context. The example preserves Engine's tested 131072 server context; it is not a claim that every runtime or context workload fits unchanged.

License

The Swift contribution remains subject to Swift Open License v1.0, including its commercial-use condition at US$1,000,000 gross annual revenue as defined in that license. Qwen base portions retain Apache 2.0. This is not an Apache-only release. Read LICENSE, LICENSE-APACHE-2.0 and NOTICE before use or redistribution. Dataset attribution does not relicense underlying repository code.

Downloads last month: 54

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for automatedigital/Paradigm-2.0

automatedigital/engine-swift-1.5-27b-agentic-lora

ukisai/Swift-1.5-Qwen3.8-27b

Merge model

this model