Instructions to use hxbibb19/Llama-Reflex-1B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use hxbibb19/Llama-Reflex-1B-GGUF 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 hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf hxbibb19/Llama-Reflex-1B-GGUF: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 hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use hxbibb19/Llama-Reflex-1B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "hxbibb19/Llama-Reflex-1B-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hxbibb19/Llama-Reflex-1B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Ollama
How to use hxbibb19/Llama-Reflex-1B-GGUF with Ollama:
```
ollama run hf.co/hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M
```
Unsloth Desktop

How to use hxbibb19/Llama-Reflex-1B-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hxbibb19/Llama-Reflex-1B-GGUF: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": "hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use hxbibb19/Llama-Reflex-1B-GGUF with Docker Model Runner:
```
docker model run hf.co/hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M
```

Lemonade

How to use hxbibb19/Llama-Reflex-1B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Llama-Reflex-1B-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use hxbibb19/Llama-Reflex-1B-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hxbibb19/Llama-Reflex-1B-GGUF: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 hxbibb19/Llama-Reflex-1B-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use hxbibb19/Llama-Reflex-1B-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hxbibb19/Llama-Reflex-1B-GGUF: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 "hxbibb19/Llama-Reflex-1B-GGUF: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"

Llama-Reflex-1B (GGUF) — Distilled from Muse Glimmer 30B

Built on Llama 3.2 1B Instruct and fine-tuned with 892 Muse Glimmer 30B teacher-labeled decisions.

Llama-Reflex-1B is the on-device "System 1" model in OffLlama, an iPhone AI assistant that keeps working without a network connection. It makes fast, typed decisions in a single forward pass: should this request go to a tool, should this fact be remembered, can a small model answer this offline, should the app confirm before acting. It accepts its own choice only when it is confident enough; otherwise the decision takes a defined fallback.

It is Llama 3.2 1B Instruct with a LoRA distilled from 892 Muse Glimmer 30B teacher labels, merged and quantized to Q4_K_M for llama.cpp.

Files

File	Size	SHA-256
`Llama-Reflex-1B-Q4_K_M.gguf`	807,694,080 bytes	`a5c60a8836aa73adeb847185aa30184cd9c33c522185f59fa9ac55a9d00782c2`
`calibration.json`	per-decision priors and the default threshold (τ = 0.90) shipped in the app	`b9ce25b5223103944803c2668784c484ac7eee75b33aa984e62185728970a8f1`

The GGUF is byte-identical to the file bundled in the OffLlama iOS app and to the GitHub release (which also carries the training labels).

How it is used

Each decision is a multiple-choice question in the Llama 3 chat format. The model is not asked to generate text; the app reads the probabilities of the option letters (" A" … " H", each a single Llama 3 token) after Answer:.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are Reflex, the fast decision module of an offline personal assistant. Answer each multiple-choice question with a single letter.<|eot_id|><|start_header_id|>user<|end_header_id|>

Context:
User: what's 2+2?

Question: Can a small on-device model answer this correctly without the internet?
A. Yes: greeting, translation of a word, small arithmetic, spelling
B. No: needs reasoning, research, long or expert output
Answer with one letter.<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Answer:

With llama-server, request one token and the top letter probabilities:

llama-server -m Llama-Reflex-1B-Q4_K_M.gguf --port 8088
curl -s localhost:8088/completion -d '{"prompt": "<prompt above>", "n_predict": 1, "n_probs": 8, "temperature": 0}'

If the top letter's probability is at least 0.90, OffLlama accepts it. Below that, it uses the decision's fallback (for example "ask the user" or "not answerable offline"). The decision types are route, toolChoice, rememberOrNot, memoryConflict, notificationTriage, confirmBeforeAct, nudgeNow, timing, sensitivity, answerability, contactPick and eventPick. The exact option sets and the Swift runtime are in the OffLlama repository.

Training

Base: Llama 3.2 1B Instruct.
Labels: 892 typed decision questions answered by Muse Glimmer 30B (Apache-2.0).
LoRA r=16, one epoch, loss on the answer letter only. bf16, effective batch 8, 112 steps.
Merged into the base model and quantized to Q4_K_M with llama.cpp.

A second epoch, other quantizations (f16, Q8_0, Q3_K_M, Q2_K) and a Muse Spark Open–labeled variant were evaluated and not shipped.

Results

On a saved 120-question decision test (10 per original decision type), using the same prompts for both models:

	Llama 3.2 1B Instruct	Llama-Reflex-1B
Correct option chosen	89 / 120 (74.2%)	98 / 120 (81.7%)
Expected calibration error (lower is better)	0.150	0.064

At the shipping threshold of 0.90, Llama-Reflex-1B accepted 44 of 120 decisions, and 41 of those 44 were correct. The other 76 took their fallback.

Limitations

This is a small test set of authored examples. It measures individual decisions, not end-to-end phone task success.
The 0.90 threshold was chosen on data that was also used for model selection, so it is not a certified error bound. See evaluation limits.
It is trained for OffLlama's fixed decision formats. It is not a general chat model, and free-form generation quality is that of a 1B model.
English only.

License

Llama-Reflex-1B is a derivative of Llama 3.2 and is distributed under the Llama 3.2 Community License. Use must follow the Llama 3.2 Acceptable Use Policy. See NOTICE.

Built by Sadman Habib for the Meta Global AI Developer Hackathon.

Downloads last month: 224

GGUF

Model size

1B params

Architecture

llama

Hardware compatibility

4-bit

Model tree for hxbibb19/Llama-Reflex-1B-GGUF

Base model

meta-llama/Llama-3.2-1B-Instruct

Finetuned

(1,850)

this model