Instructions to use Thox-ai/ThoxClip-9M-role with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Thox-ai/ThoxClip-9M-role 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 Thox-ai/ThoxClip-9M-role:TQ1_0
# Run inference directly in the terminal:
llama cli -hf Thox-ai/ThoxClip-9M-role:TQ1_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Thox-ai/ThoxClip-9M-role:TQ1_0
# Run inference directly in the terminal:
llama cli -hf Thox-ai/ThoxClip-9M-role:TQ1_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 Thox-ai/ThoxClip-9M-role:TQ1_0
# Run inference directly in the terminal:
./llama-cli -hf Thox-ai/ThoxClip-9M-role:TQ1_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 Thox-ai/ThoxClip-9M-role:TQ1_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Thox-ai/ThoxClip-9M-role:TQ1_0

Use Docker

docker model run hf.co/Thox-ai/ThoxClip-9M-role:TQ1_0

LM Studio
Jan

vLLM

How to use Thox-ai/ThoxClip-9M-role with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Thox-ai/ThoxClip-9M-role"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Thox-ai/ThoxClip-9M-role",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Thox-ai/ThoxClip-9M-role:TQ1_0

Ollama
How to use Thox-ai/ThoxClip-9M-role with Ollama:
```
ollama run hf.co/Thox-ai/ThoxClip-9M-role:TQ1_0
```

Unsloth Studio

How to use Thox-ai/ThoxClip-9M-role with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Thox-ai/ThoxClip-9M-role to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Thox-ai/ThoxClip-9M-role to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Thox-ai/ThoxClip-9M-role to start chatting

Docker Model Runner
How to use Thox-ai/ThoxClip-9M-role with Docker Model Runner:
```
docker model run hf.co/Thox-ai/ThoxClip-9M-role:TQ1_0
```

Lemonade

How to use Thox-ai/ThoxClip-9M-role with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Thox-ai/ThoxClip-9M-role:TQ1_0

Run and chat with the model

lemonade run user.ThoxClip-9M-role-TQ1_0

List all available models

lemonade list

Atomic Chat

ThoxClip-9M-role

On-device model for ThoxClip (Pi Zero 2 W, arm64/NEON). BitNet b1.58 ternary, 8,917,248 params, 76.4% of weights in {-1, 0, +1}.

Trained for the device role: handle local commands, name itself when asked, and defer everything else rather than invent it.

Why this exists separately from ThoxKey-9M-role

The first ThoxClip drop reused ThoxKey's corpus verbatim, which produced a Pi Zero that answered "No. I run on the key." Correct in substance, wrong device. This is retrained on ThoxClip's own corpus so the nouns and identity are right.

Files

file	bytes
`thoxclip-9m-role-TQ1_0.gguf`	5,933,152
`thoxclip-9m-role.tern1`	4,420,028

GGUF export verified near-lossless: 56 tensors, worst relative error 4.72e-04 against a 2e-03 tolerance.

Ollama cannot load ternary GGUF (TQ1_0/TQ2_0/I2_S) — it fails with tensor "blk.0.ffn_down.weight" size overflow. Use llama.cpp.

Training


init	`Thox-ai/ThoxMicro-1bit-9M` (393M-token pretrain)
corpus	25% ThoxClip device-role / 75% TinyStories, 30M tokens
steps	1,500 × 16,384 tok = 24.6M tokens
hardware	local RTX 4060 Ti, 330 s, $0
val loss	1.8001 (pretrain baseline 1.8281)

Validation is pure TinyStories, excluding the device corpus, so the number cannot be flattered by memorising device turns.

Status

Published ship-first. On-device performance is not yet measured on Pi Zero 2 W hardware — post-implementation testing and hotfixes follow integration.

Downloads last month: 12

GGUF

Model size

8.92M params

Architecture

llama

Hardware compatibility

1-bit

Model tree for Thox-ai/ThoxClip-9M-role

Base model

Thox-ai/ThoxMicro-1bit-9M

Quantized

(2)

this model