Open beta. Registered accounts may submit free browser scans of eligible public model revisions through the 7B class, within their account quota. Models above the 7B class are not scanned in this beta. Results remain research-stage. Payment collection and invoicing are disabled.
Compare model checkpoints, understand the recorded changes, and choose the next checks for your own workload. Model X-Ray adds evidence to your evaluation process; it does not decide whether a model is better or ready to ship.
Your account page shows whether you can submit a scan and how much allowance remains.
The standard free beta allowance is 20 scans per month for eligible public models through the 7B class. Your account may have a different owner-assigned allowance. Existing Pro or Team plans stay unchanged; a separate free beta grant can provide browser scans without changing paid quotas or payment history. Paid activation and payment collection remain paused.
organization/model form. The service records exact repository revisions before accepting the job.A checkpoint is a saved set of model weights. A revision identifies the exact repository version used. A comparison with different configuration, tokenizer or generation settings is a combined artifact change; do not attribute it to fine-tuning alone.
| Scan | Inputs | Useful for |
|---|---|---|
| Compare two checkpoints | Supported reference A and candidate B | Recording internal-response and eligible output-text differences, with artifact context and the report's available depth detail. |
| Single X-Ray: quantization preview | One source checkpoint and an int8 or int4 simulation | Exploring a lower-precision state before testing an actual deployable quantized artifact. |
A simulation is not a deployed quantized model. Its B state is created in memory from A and has no separate repository revision or downloadable model file. It does not measure packed-kernel speed, memory savings or task quality. Test those on the real artifact and inference implementation.
| Section | How to read it |
|---|---|
| Internal response | “Difference observed” means the recorded comparison met the versioned measurement rule on the evaluated inputs. “No difference observed” does not establish equivalence on other inputs or tasks. |
| Text output | Changed or unchanged refers to exact generated token sequences under a comparable output contract. “Withheld” means the output comparison was not eligible; it does not mean unchanged. None of these is a semantic-quality score. |
| Depth view | When included, the categorical map and its text table mark evaluated positions as difference observed, no difference observed, or not evaluated. Read the boundary note before interpreting the first recorded difference. |
| Verify view | This view presents the finding, output status and available verification summaries without internal-position detail. It does not change the measured finding. |
| Artifact context | The report distinguishes matched and different configuration, tokenizer and generation summaries. Matched summaries do not prove that complete artifacts are identical. |
| Measured weight changes | Only present when separately validated weight evidence is included. Model and parameter-group summaries describe recorded weight changes alongside response observations; larger values do not mean greater harm or better quality. |
| Included summaries | Check whether repeat measurements, same-artifact controls and an operational receipt are actually included. “Not included” is not a pass or a failed control. |
Depth is a recorded response, not the location of an edit. A difference can propagate to later outputs. If it is already present at the first measured point, the beginning is unresolved. “Not evaluated” is unknown, not a below-threshold result. Final denotes the output after the final normalization; it is not a separate measurement of the last block before normalization.
Use a depth view only within that report's stated measurement and eligibility limits. Controlled findings on another model do not validate a causal explanation for yours. The report does not identify edited weights from response depth, diagnose a capability, or provide a calibrated severity or pass/fail grade.
Repeat and control summaries describe only their stated executions. Matching fresh-process measurements within one allocation do not prove agreement across GPUs, machines or allocations. No model-specific statistical noise envelope is implied by a fixed decision rule.
For continued training, compare later checkpoints against a consistent reference and keep your training change log with the reports. The scan is not an optimization target, a model ranking, or a replacement for your acceptance tests.
Owner HTML: the dashboard report and HTML download provide the customer report, including its available categorical depth and optional weight summaries. Download retains the original single-file HTML. It has no scripts or remote assets; verification links connect only when followed.
JSON summary: the dashboard's JSON link at /report/<job-id>/summary.json is deliberately smaller. It contains the curated observation, eligible output-token status and record context. It is not an export of the full HTML depth or weight tables, raw measurements, private traces or a calibrated quality grade.
Records: a signature authenticates the stated Tetracta operational record. It is not independent proof of task quality, hardware identity or physical deletion. Some review copies contain no signed receipt; their included-evidence section says so.
Report sharing and headless API scan access are currently unavailable through free beta. There is no separate PDF download endpoint. Opening or downloading an accepted owner report does not use another scan.
Standard beta defaults to 20 scans per month and five model slots, through the 7B class. A slot tracks a distinct model with retained jobs in the current allowance period. Removing a report can free its model slot when no other retained job uses it; it does not refund a completed scan. The account page shows your applicable allowance. A one-time trial, when assigned, has 1 scan.
A historical paid plan does not automatically enable free scans or paid scanning. A separate owner-issued free beta allowance has its own usage accounting; it preserves the original plan, paid quota and payment history. Model eligibility, allowance and available capacity are checked again when a job is submitted. Models above the 7B class are not included in this beta.
Browser scans accept eligible public Hugging Face safetensors artifacts that fit the supported model interface. A repository name alone does not guarantee support. Private tokens, customer uploads, Training Watch and free API submission are unavailable. Structural Portrait and Knowledge scans remain unavailable; historical claims and corrections retain their original scope at the correction record.
Which repositories, which files, how long. The eligible repositories are the ones listed on the scope page; the new-scan form offers a few of them as suggestions, and the exact revision is fixed by the rights register and recorded in the report. A scan reads safetensors weights under their canonical names (model.safetensors, or model.safetensors.index.json with its shards); a repository that publishes only a pickle checkpoint or only another format such as GGUF is declined before a job is created and does not use your allowance. Quantized release formats (FP8, GPTQ, AWQ, GGUF) are not measured; the quantization simulation is derived in memory from the full-precision weights. Both checkpoints in a comparison must share the same architecture. A comparison keeps both snapshots on one worker and loads them one after the other; a pair of two 7B-class checkpoints is currently accepted by the form but fails inside the worker sandbox (under investigation, 22 September 2026), so please do not submit one until this sentence is removed. A single 7B-class checkpoint runs normally as a quantization simulation. Timing: a 7B-class scan has taken between about 8 and 12 minutes end to end in the three recorded runs on the current 7B-capable worker. Queue time is shown on the dashboard and is not included in these figures.
Need help? Write to hello@tetracta.ai with the visible error and scan reference. Do not send passwords or access tokens.
Model X-Ray reports are descriptive research evidence. No hallucination, task-accuracy, safety, overtraining, model-superiority or deployment-readiness certification is provided. Keep the report's honesty box with any copy you share outside the application.
Tetracta · pre-release validation · public-reference weights follow the managed worker cache policy; any future private/upload release requires a separate reviewed deletion boundary · deleting a report removes the live copy immediately; disaster-recovery backup retention and deletion-after-restore controls remain under review, with no fixed expiry promised in this pre-release · docs · security · terms · privacy · software notices