Instructions to use mburchell/looped-40m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mburchell/looped-40m with Transformers:

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

pipe = pipeline("text-generation", model="mburchell/looped-40m", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("mburchell/looped-40m", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mburchell/looped-40m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mburchell/looped-40m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mburchell/looped-40m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mburchell/looped-40m

SGLang

How to use mburchell/looped-40m 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 "mburchell/looped-40m" \
    --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": "mburchell/looped-40m",
		"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 "mburchell/looped-40m" \
        --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": "mburchell/looped-40m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mburchell/looped-40m with Docker Model Runner:
```
docker model run hf.co/mburchell/looped-40m
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Overview

A from-scratch experimental 40M Llama-style model utilizing recurrence. 512-wide, SwiGLU, GQA, 10 layers. Loops once across the inner 6 layers. 1024 context. 24K custom vocab with case operators.

Trained on approx. 15B total tokens using Muon.

Trained for light chat with Llama-3 template. It's relatively cohesive but starts to fall apart as context grows. It is likely biased, has no guardrails, and is completely useless for any real tasks.

Example conversation:

prompt> What is machine learning?

Machine learning is a branch of computer science that aims to identify patterns in data or patterns from multiple sources. The aim of machine learning is to apply new knowledge or techniques to improve or replace existing methods.

prompt> Can machine learning apply to language?

Yes, it can apply to language, where it helps in understanding and interpreting natural language. It's also possible to learn languages by watching videos or analyzing the results from real-world examples.

prompt> Are you a machine learning language model?

Yes, machine learning models are used in many fields, such as data science, natural language processing, and information technology. They can analyze large datasets or perform complex tasks, such as querying and coding languages.

Training sources, attribution and publication

The following section is AI generated from assistant session summaries.

See provenance.json for pinned dataset revisions, hashes and token counts.

The pretraining source is the FineWeb-Edu-dedup subset of HuggingFaceTB/SmolLM-Corpus, published under ODC-BY; it includes educational web text with underlying rights.

Conversational SFT used:

SQuAD 1.1, by Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev and Percy Liang: CC BY-SA 4.0, with Wikipedia-derived passages.
Smol-SmolTalk, by Hugging Face: Apache 2.0; selected short, constraints, rewriting and summarization examples.
Everyday Conversations, by Hugging Face: Apache 2.0.
bAbI QA, by Jason Weston and colleagues/Facebook AI Research: CC BY 3.0; selected tasks.

The finishing run mixed 60% local synthetic chats, 20% Everyday and 20% multi-turn Smol-SmolTalk. Synthetic replies came from a locally served gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf, with programmatic factual answers, templates, edits and validation.

Downloads last month: -

Safetensors

Model size

40M params

Tensor type

F32

Model tree for mburchell/looped-40m

Quantizations

1 model

mburchell
/

looped-40m

Overview

Example conversation:

Training sources, attribution and publication

Model tree for mburchell/looped-40m

Datasets used to train mburchell/looped-40m