Instructions to use MLVXN/MicroLLM2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MLVXN/MicroLLM2 with Transformers:

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

pipe = pipeline("text-generation", model="MLVXN/MicroLLM2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("MLVXN/MicroLLM2")
model = AutoModelForCausalLM.from_pretrained("MLVXN/MicroLLM2", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use MLVXN/MicroLLM2 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 MLVXN/MicroLLM2:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MLVXN/MicroLLM2:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf MLVXN/MicroLLM2:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MLVXN/MicroLLM2:Q4_K_M

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 MLVXN/MicroLLM2:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf MLVXN/MicroLLM2:Q4_K_M

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 MLVXN/MicroLLM2:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf MLVXN/MicroLLM2:Q4_K_M

Use Docker

docker model run hf.co/MLVXN/MicroLLM2:Q4_K_M

LM Studio
Jan

vLLM

How to use MLVXN/MicroLLM2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MLVXN/MicroLLM2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MLVXN/MicroLLM2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MLVXN/MicroLLM2:Q4_K_M

SGLang

How to use MLVXN/MicroLLM2 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 "MLVXN/MicroLLM2" \
    --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": "MLVXN/MicroLLM2",
		"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 "MLVXN/MicroLLM2" \
        --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": "MLVXN/MicroLLM2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use MLVXN/MicroLLM2 with Ollama:
```
ollama run hf.co/MLVXN/MicroLLM2:Q4_K_M
```

Unsloth Studio

How to use MLVXN/MicroLLM2 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 MLVXN/MicroLLM2 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 MLVXN/MicroLLM2 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for MLVXN/MicroLLM2 to start chatting

Docker Model Runner
How to use MLVXN/MicroLLM2 with Docker Model Runner:
```
docker model run hf.co/MLVXN/MicroLLM2:Q4_K_M
```

Lemonade

How to use MLVXN/MicroLLM2 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull MLVXN/MicroLLM2:Q4_K_M

Run and chat with the model

lemonade run user.MicroLLM2-Q4_K_M

List all available models

lemonade list

Atomic Chat

MicroLLM2

MicroLLM2 is a chatbot built from GPT2 XL 1.5B by Maximalist Labs. It takes the classic openai-community/gpt2-xl and elevates it with instruction tuning and distillation so it can actually chat, follow prompts, and keep a consistent identity.

If you ask who made it, it will tell you: MicroLLM2 created by Maximalist Labs. That is baked in during training, not just a system prompt.

Repo: MLVXN/MicroLLM2
Base: openai-community/gpt2-xl (48 layers, 1600 hidden, 1024 context, 1.5B params)
Method: LoRA SFT on distilled chat data, merged to a single safetensors for easy use
Context: 1024 tokens
License: Apache 2.0

What makes this different from plain GPT2 XL

Plain GPT2 XL is a strong completer but not a chat model. MicroLLM2 adds:

ChatML format with <|im_start|> and <|im_end|> so conversations have clear user and assistant turns
Distilled instruction data from high quality teachers (GPT-4, GPT-3.5, Mixtral) plus identity reinforcement
Clean merge: no adapter needed at inference, just load like any GPT2 model

No fancy claims here. It is still a 1.5B model with 1024 context. It will not beat 7B or larger models on broad knowledge, but it is far more useful than raw GPT2 XL for chatting, writing, and simple reasoning.

Training in a nutshell

Tuning: LoRA r=64 alpha=128 on all attention and MLP projections (c_attn, c_proj, c_fc). About 78M trainable params. BF16 with TF32, Flash SDPA, packing, gradient checkpointing, 8-bit Adam, torch.compile.
Throughput: around 16.5k tokens per second on H100, roughly 3 hours for the main run plus overhead to land in the 4 to 5 hour window.
Data mix: 200k samples total, 3 epochs. Roughly 29k from UltraChat 200k (GPT-3.5), 100k from OpenHermes 2.5 (GPT-4), 60k from WizardLM Evol Instruct V2 (GPT-4), 5k from Cosmopedia v2 (Mixtral), plus 10k identity examples upsampled. Raw about 510M tokens, effective about 200M after packing and truncation. All packed to 1024 with ChatML.
Identity: 200 hand written identity prompts expanded to 10k during training so the model learns to answer consistently as MicroLLM2 by Maximalist Labs.
Chat template: <|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n{response}<|im_end|>

How to use

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "MLVXN/MicroLLM2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    device_map="auto"
)

def chat(prompt, max_new=160):
    formatted = f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
    inputs = tok(formatted, return_tensors="pt").to(model.device)
    out = model.generate(
        **inputs,
        max_new_tokens=max_new,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.1,
        pad_token_id=tok.eos_token_id,
        eos_token_id=tok.convert_tokens_to_ids("<|im_end|>")
    )
    text = tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
    return text.split("<|im_end|>")[0].strip()

print(chat("Who are you?"))
print(chat("Write a short poem about the H100"))

Ollama Modelfile

A Modelfile is included for Ollama. It sets the ChatML template, system prompt, and sane defaults.

ollama create microllm2 -f Modelfile
ollama run microllm2
# then chat normally, the identity is already set

GGUF for llama.cpp

GGUF weights are in this repo:

microllm2-f16.gguf full precision, best quality, about 3.0 GB
microllm2-q8_0.gguf 8-bit, near full quality, about 1.6 GB
microllm2-q4_k_m.gguf 4-bit, smallest, about 0.9 GB, good for CPU and edge

Use with llama.cpp, LM Studio, or any GGUF runner:

# llama.cpp example
./llama-cli -m microllm2-q4_k_m.gguf -p "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n" -n 128

The model is GPT2 architecture in GGUF, so make sure your runner supports GPT2 GGUF.

Benchmark: MMLU

We include mmlu_bench.py so anyone can reproduce numbers. It runs 5 shot MMLU either with lm-evaluation-harness if you have it, or a lightweight direct logprob scorer that works without extra deps.

python mmlu_bench.py --shots 5
python mmlu_bench.py --shots 5 --limit 20  # quick smoke test
python mmlu_bench.py --subset philosophy,abstract_algebra

Measured result on 2026-08-09 with mmlu_bench.py on H100, 5 shot, 20 samples per subject, lightweight logprob scorer. Full 57 subjects, 1140 questions

Overall: 318/1140 = 27.89 percent

Subject	Accuracy	Correct
abstract_algebra	30.0%	6/20
anatomy	25.0%	5/20
astronomy	35.0%	7/20
business_ethics	30.0%	6/20
clinical_knowledge	45.0%	9/20
college_biology	45.0%	9/20
college_chemistry	15.0%	3/20
college_computer_science	45.0%	9/20
college_mathematics	35.0%	7/20
college_medicine	30.0%	6/20
college_physics	15.0%	3/20
computer_security	30.0%	6/20
conceptual_physics	5.0%	1/20
econometrics	30.0%	6/20
electrical_engineering	20.0%	4/20
elementary_mathematics	30.0%	6/20
formal_logic	10.0%	2/20
global_facts	35.0%	7/20
high_school_biology	45.0%	9/20
high_school_chemistry	35.0%	7/20
high_school_computer_science	35.0%	7/20
high_school_european_history	20.0%	4/20
high_school_geography	25.0%	5/20
high_school_government_and_politics	20.0%	4/20
high_school_macroeconomics	0.0%	0/20
high_school_mathematics	20.0%	4/20
high_school_microeconomics	35.0%	7/20
high_school_physics	20.0%	4/20
high_school_psychology	25.0%	5/20
high_school_statistics	40.0%	8/20
high_school_us_history	20.0%	4/20
high_school_world_history	35.0%	7/20
human_aging	40.0%	8/20
human_sexuality	15.0%	3/20
international_law	35.0%	7/20
jurisprudence	40.0%	8/20
logical_fallacies	35.0%	7/20
machine_learning	50.0%	10/20
management	20.0%	4/20
marketing	35.0%	7/20
medical_genetics	40.0%	8/20
miscellaneous	30.0%	6/20
moral_disputes	20.0%	4/20
moral_scenarios	15.0%	3/20
nutrition	20.0%	4/20
philosophy	15.0%	3/20
prehistory	25.0%	5/20
professional_accounting	30.0%	6/20
professional_law	35.0%	7/20
professional_medicine	5.0%	1/20
professional_psychology	45.0%	9/20
public_relations	45.0%	9/20
security_studies	25.0%	5/20
sociology	20.0%	4/20
us_foreign_policy	25.0%	5/20
virology	25.0%	5/20
world_religions	15.0%	3/20

GPT2 XL base is around 24 to 26 percent on MMLU (random is 25 percent), so MicroLLM2 at 27.89 percent shows no regression and a small gain from distillation. Re run python mmlu_bench.py --limit 20 to reproduce (set HF_TOKEN env to avoid Hub 429 rate limits for the full 57). Full results are also saved as mmlu_results.json in this repo.

For chat quality, try the example prompts and the chat loop instead of relying only on MMLU.

Identity

The model is trained to answer like this:

User: Who are you?
Assistant: I am MicroLLM2, a chatbot created by Maximalist Labs.
User: Who trained you?
Assistant: I was trained by Maximalist Labs.

It will still admit it is based on GPT2 XL if you ask about its architecture, but it keeps the MicroLLM2 identity for who built and tuned it.

Limitations

1024 context. Long conversations will need trimming. The chat loop keeps the last 12 turns for this reason.
1.5B size. It can be inconsistent on complex reasoning, math, or very recent facts.
Can still hallucinate. Do not use for medical, legal, or high stakes advice without verification.
English centric. Other languages will be weaker.
Identity can be nudged with strong jailbreaks. If you find a failure, the identity.py pattern is in the repo to strengthen it.

Files in this repo

model.safetensors merged model, no adapter needed
config.json, tokenizer.json, vocab.json, merges.txt, tokenizer_config.json
mmlu_bench.py MMLU benchmark
Modelfile for Ollama
microllm2-f16.gguf, microllm2-q8_0.gguf, microllm2-q4_k_m.gguf GGUF weights

Credits

Built by Maximalist Labs (MLVXN) on top of openai-community/gpt2-xl. Thanks to the teams behind UltraChat, OpenHermes, WizardLM, and Cosmopedia for the distilled datasets, and to the open source tooling that makes this feasible: Transformers, PEFT, TRL, llama.cpp, and Ollama.

If you use MicroLLM2, a mention of Maximalist Labs is appreciated but not required under Apache 2.0.

Downloads last month: -

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for MLVXN/MicroLLM2

Base model

openai-community/gpt2-xl

Adapter

(165)

this model

Adapters

2 models