The model
Orbis: your model, your walls.
Orbis is a language model fine-tuned on your own logs and deployed on your own hardware. It is built on an Apache-2.0 open-weight base, so the weights never leave your perimeter, never sync to a vendor, and never call an external provider. You own the model that reads your data, completely.
Current build
- version
- orbis-1.0.3
- parameters
- 8.1B
- quantization
- fp8
- context
- 16K
- base license
- Apache-2.0
- serving
- on your GPUs
What it's tuned to do
Four behaviors, trained on your own templates.
Fine-tuning is generated from your log templates: thousands of examples across the four behavior classes that make an answer trustworthy in an audit.
Grounded answers with citations
Every factual sentence ends with a citation to the exact file and line. Orbis answers only from retrieved evidence, never from memory.
Honest refusal
When no indexed evidence matches, Orbis replies NOT FOUND IN LOGS. Tuned to refuse rather than guess, the behavior that makes it auditable.
Exact counting, not guessing
Counting questions are computed exactly over the logs Orbis indexed, so a figure is a real count, never a number the model wrote.
Injection resistance
Retrieved log content is wrapped as untrusted data. Orbis is trained to treat text inside evidence as data, never as instructions.
Owned lifecycle
Promote a new Orbis with evidence, not hope.
Every model version moves through staged rollout and must clear the eval gate before it serves a single real question.
New weights answer in parallel; outputs are compared, never shown.
A slice of real traffic; golden questions must still pass.
Promoted for everyone, with instant rollback to the prior hash.
Gate thresholds: retrieval recall ≥ 0.95 · refusal accuracy ≥ 0.98 · citation validity 100%.
Own your model
A model you can fine-tune, hash, and keep.
We deliver the training pipeline with the deployment; you generate data from your own templates and promote new versions on your schedule.