Open language models, compressed to fit in about 1 GB.

An independent project that measures how far a language model can be shrunk before quality drops, and what each cut costs. No lab, on a single 8 GB consumer graphics card.

69% smaller than the original, on the 1.5-billion-parameter model

92% of the original's general-knowledge score, kept

0.90 GiB in the GGUF format, the same phone apps use, offline

Open base model Qwen2.5-1.5B (Alibaba). Sizes in GiB (the unit Windows shows as GB), measured on the real file. Score on the MMLU test with 1,000 questions, on the same yardstick for every model, measured before the conversion to GGUF; the conversion was checked in llama.cpp.

~100models trained from scratch in the first phase

5open models evaluated in the current phase

1 billiontext tokens prepared for training

5,000questions in the largest knowledge test

1consumer GPU, 8 GB

WHAT CHANGED

The project's first phase tested a new architecture that reuses the same processing blocks several times. It shrank the model and kept almost all of its quality, but a stricter control showed that an ordinary model of the same size ties or wins. The thesis was retracted.

Since then the question changed: instead of inventing a smaller model, shrink open models that already exist until they fit in about 1 GB, and measure rigorously how much quality is left at each size.

THE RESULTS

How much each model shrank

Size of the original model (light bar, 16 bits per parameter) and of the compressed version (amber bar, measured on the file written to disk), in GiB.

How much quality was left

Score on the MMLU general-knowledge test (1,000 questions), as a percentage of the original model's score. The more a model is shrunk relative to its original, the more score it loses. On a 4-option test, guessing already scores 25%.

HOW IT WAS TESTED

Prediction registered before the experiment. The training tests and the decisive test had the expected result and the success criterion written down before running. Predictions that failed are recorded as failures.
Size in real bytes. Each compressed version's size is measured on the whole file written to disk, not just the convenient part.
The same yardstick for everyone. The compressed models and the originals go through the same evaluator, on the same questions.
Question-by-question comparison. In the decisive comparisons, two versions are only declared different if the difference is statistically significant. 200-question samples were dropped because they were off by more than 5 points.
Audit before running. Every script is reviewed before it runs, and every result is checked again against the files before it is published.

LIMITS OF WHAT THIS SHOWS

The base models are open and third-party (Qwen, by Alibaba). The project's work is the compression, the training and the measurement.
The measured gain is in file size. Running faster or with less memory is not a claim of this work.
The 3-billion model at about 1 GB still loses a lot: it keeps 74% of the score and gets math and code wrong. It is research, not a product.
Each training run was done once, without repetition.
Only Qwen-family models were tested, and quality is measured with standard tests, which do not cover everything a model does.