Q Project
Independent AI research lab
Q Project trains language and agent models from scratch — some open for anyone to run, some kept private. Every weight, tokenizer, and training decision is ours.
Open-weight releases to learn from, and closed models built to go further.
A 50.9M-parameter decoder-only language model, pretrained from scratch on 5 billion tokens of FineWeb-Edu. A base model for continuation, study, and fine-tuning — small enough to run anywhere.
Our next model is on the cluster now — a larger system built for perception and for acting through tools, not just predicting text. Details stay under wraps until it is ready.
It loads with the standard Transformers API. No sign-up, no gate — the weights are open.
Install the library, pull the checkpoint from the Hub, and let it continue your text.
Full quickstart guideQ-50M-Base continues text rather than following instructions. For chat behavior, fine-tune it on your own data.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "q-project/Q-50M-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16)
prompt = "The theory of relativity states that"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=64, do_sample=True, temperature=0.8)
print(tokenizer.decode(output[0], skip_special_tokens=True))Four threads that feed each other, from raw pretraining to models that take actions.
Training foundations from scratch — our own tokenizers, data mixes, and stable, efficient architectures.
Turning base models into capable, well-behaved ones through supervised and preference tuning.
Models that use tools, perceive their surroundings, and carry out multi-step tasks reliably.
Publishing what we learn — models, datasets, and notes on making small models punch above their size.
Open weights stay free. Hosted access and private models are on the way — join the waitlist and we’ll let you know.
Freealways
For builders, students, and researchers.
Coming soonpricing at launch
Hosted API for Q models, without running your own hardware.
Let’s talk
Closed models, custom training, and dedicated support.