Instructions to use NotHereNorThere/AtomCoT-135M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use NotHereNorThere/AtomCoT-135M with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="NotHereNorThere/AtomCoT-135M")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("NotHereNorThere/AtomCoT-135M", dtype="auto")

llama-cpp-python

How to use NotHereNorThere/AtomCoT-135M with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="NotHereNorThere/AtomCoT-135M",
	filename="model-Q6_K.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use NotHereNorThere/AtomCoT-135M with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf NotHereNorThere/AtomCoT-135M:Q6_K
# Run inference directly in the terminal:
llama-cli -hf NotHereNorThere/AtomCoT-135M:Q6_K

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf NotHereNorThere/AtomCoT-135M:Q6_K
# Run inference directly in the terminal:
llama-cli -hf NotHereNorThere/AtomCoT-135M:Q6_K

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 NotHereNorThere/AtomCoT-135M:Q6_K
# Run inference directly in the terminal:
./llama-cli -hf NotHereNorThere/AtomCoT-135M:Q6_K

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 NotHereNorThere/AtomCoT-135M:Q6_K
# Run inference directly in the terminal:
./build/bin/llama-cli -hf NotHereNorThere/AtomCoT-135M:Q6_K

Use Docker

docker model run hf.co/NotHereNorThere/AtomCoT-135M:Q6_K

LM Studio
Jan

vLLM

How to use NotHereNorThere/AtomCoT-135M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "NotHereNorThere/AtomCoT-135M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotHereNorThere/AtomCoT-135M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/NotHereNorThere/AtomCoT-135M:Q6_K

SGLang

How to use NotHereNorThere/AtomCoT-135M with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "NotHereNorThere/AtomCoT-135M" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotHereNorThere/AtomCoT-135M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "NotHereNorThere/AtomCoT-135M" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotHereNorThere/AtomCoT-135M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use NotHereNorThere/AtomCoT-135M with Ollama:
```
ollama run hf.co/NotHereNorThere/AtomCoT-135M:Q6_K
```

Unsloth Studio

How to use NotHereNorThere/AtomCoT-135M 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 NotHereNorThere/AtomCoT-135M 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 NotHereNorThere/AtomCoT-135M to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for NotHereNorThere/AtomCoT-135M to start chatting

Docker Model Runner
How to use NotHereNorThere/AtomCoT-135M with Docker Model Runner:
```
docker model run hf.co/NotHereNorThere/AtomCoT-135M:Q6_K
```

Lemonade

How to use NotHereNorThere/AtomCoT-135M with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull NotHereNorThere/AtomCoT-135M:Q6_K

Run and chat with the model

lemonade run user.AtomCoT-135M-Q6_K

List all available models

lemonade list

AtomCoT-135M

SmolLM2-135M-Instruct fine-tuned on OpenThoughts data for chain-of-thought reasoning.

The smallest possible unit of attempted CoT transfer. Emphasis on attempted.

What it is

QLoRA fine-tune of SmolLM2-135M-Instruct exploring the lower bound of coherent reasoning behavior. If Qwen3 600M was surprisingly good for its size, AtomCoT-135M is the control group that explains why 600M is surprising.

Training

Setting	Value
Base model	`HuggingFaceTB/SmolLM2-135M-Instruct`
Method	QLoRA (4-bit NF4, LoRA r=16)
Dataset	OpenThoughts 2k subset
Hardware	RTX 4060 8GB
Attention	FlashAttention 2

Eval results

Prompt	Result	Notes
Bat & ball ($1.10 problem)	❌ Wrong	Invented phantom variables, concluded ball costs "0.115x the bat"
Jug problem (3gal + 5gal = 4gal)	❌ Wrong	Started measuring jug diameter in feet, concluded answer is "1 foot above the top"
6/12 fish drowning	❌ Wrong	Rewrote the question mid-answer to include the solution, then stopped
15% of 200	❌ Mega Wrong	Correctly got 0.15, then multiplied 200 by 10,000 and got 1,500,000

Honest assessment

The CoT skeleton transferred perfectly. Alost every response starts with structured steps, attempts to show work, and maintains the visual format of reasoning. The content inside that skeleton is complete chaos.

135M parameters is enough to learn what reasoning looks like but not enough to actually execute it. The model wears the costume without understanding the role. Notable failure modes include inventing intermediate steps from nowhere, conflating unrelated units and quantities, and in one case rewriting the question itself to contain the answer before giving up.

This is not a failure of fine-tuning. This is the parameter cont floor.

Key finding

The jump from 135M to 500M (tested against Qwemini-0.5B-Alpha) is where something qualitatively real switches on. AtomCoT produces structured nonsense. Qwemini produces structured correct answers on simple problems. Same training approach, same dataset style, 3.7x the parameters makes the difference between a reasoning costume and actual reasoning.