Work / Model X-Ray
Model X-Ray
Compare two checkpoints of a model, or a model and a simulated quantisation of it, under one fixed measurement setup, and get a report of the differences measured between them. For teams that fine-tune, merge or compress models. The open beta reads public models from Hugging Face by name, up to the 7B class, and scans them on our own GPUs, free within a monthly allowance. A 7B-class model runs as a quantisation simulation of that one model.
Open the beta Ask about models above the 7B class
- Status
- Open beta
- Payments
- Off during the beta
- Model size
- Up to the 7B class, since 21 Sep 2026
- Free allowance
- 20 scans a month
- Runs on
- Our own GPU workers
- Updated
- 11 Nov 2026
What runs today
- 7B-class models. These checkpoints run as single-model quantisation simulations.
- Twenty-one public repositories. Each is an exact Apache-2.0 revision with a completed rights review, listed on the scope page.
- Free within an allowance. A registered account's browser scans renew each calendar month and can cover at most five distinct source models.
- Our own hardware. Scans run on Tetracta-operated GPU workers. No cloud GPU is used in this beta.
- Safetensors weights. A repository that publishes only another format is declined before a job is created and does not use your allowance.
Models above the 7B class are not scanned in this free beta. For those, write to us and we arrange the work outside the beta.
What it does
Model X-Ray compares two sides under one fixed measurement setup:
- two checkpoints of a model, for example before and after fine-tuning, or a base model and a merge built from it;
- a model and a simulated quantisation of the same model, made in memory.
A scan puts both sides through the same measurements. It measures whether the model's internal responses differ on the evaluated inputs, whether the generated text changed, and how far each group of weights moved from the reference. The scope page explains each report item and its limit. Because the setup is fixed, two reports made the same way can be set side by side.
Using the beta
The open beta runs at tetracta.ai/llm_tomografi, and you register here. The address keeps the tool's original working name. Payments are switched off: we are not charging anyone during the beta.
Each run produces a report of measured differences between the two sides. The report format, the comparisons on offer and the service itself may change while the beta runs.
Sample reports
Static sample reports are on Hugging Face: Model X-Ray sample reports.
These show the report format only. Do not read them as findings about the models in them.
What it is not
- Diagnosis. The report lists differences. It does not certify quality, safety or performance.
- Ranking. It does not score or rank models.
- Deployed quantisation. The quantisation it compares against is simulated in memory. A model quantised for deployment, running on its own kernels, may behave differently.
- Stability of the format. The beta is still changing (see Using the beta).
Correction record
Withdrawn On 4 to 6 September 2026 we withdrew the earlier public findings we had drawn from Model X-Ray; they are not current evidence. Status on 11 November 2026: the correction is applied and the beta is open. Read the dated correction record
Compare two models
Register for the open beta and scan a public model.