Confidential AI for research and legal teams

VirexaLLM runs open-weight models locally on the attorney's or researcher's own hardware — with pinnable versions, signed audit logs, and full air-gap support. Privilege and reproducibility, enforced by the architecture.

Research moves forward without leaving the perimeter

100%

On-Device

Discovery material never leaves the workstation

Air-Gap

Supported

Run inference on machines with no network at all

Pinnable

Model Versions

Lock research to a specific open-weight revision

Signed

Audit Trail

Tamper-evident logs of every query and model load

Privilege, pinning, and provenance

Everything you need to use AI on sensitive material without exposing it.

Privileged, On-Device Inference

Contracts, filings, privileged memos, and discovery material stay on the attorney's hardware. No third-party processor to add to your engagement letter.

Attorney-Client Friendly

No vendor inspecting prompts. No training on your queries — because there's no backend to train. Confidentiality is a property of the architecture, not a checkbox.

Reproducible Research

Pin a research workflow to an exact open-weight model revision. Re-run the same query against the same weights six months later and get the same answer.

Pinned Model Versions

Open-weight models don't silently change. Freeze a model at a specific GGUF hash for regulated research, citations, and expert reports.

Signed Local Audit Trail

Every query, every model load, every policy change recorded in a tamper-evident log on the device — export to your matter management system without exposing prompt content.

Dataset Replay

Replay a historical query set against any new open-weight model with one command. Compare research outputs across model generations, reproducibly.

Work teams run on VirexaLLM

Reproducible, confidential, and fully on-device.

Discovery ReviewPrivileged Document Q&AContract AnalysisDeposition PrepCitation CheckingRegulatory ResearchConfidential SummarizationBrief DraftingAcademic ResearchReproducible ExperimentsAir-Gapped AnalysisExpert Report Support

One device, every confidential query

Load Llama 3, Mistral, Qwen, or DeepSeek directly on the researcher's machine. Every question about a privileged memo, a draft brief, or a confidential dataset gets answered by a model that never phones home.

Reproducibility, enforced by hashes

Open-weight models don't silently update. Pin a matter or a study to a specific GGUF hash and quantization. Six months later, the same query against the same weights produces the same answer — the foundation of defensible research.

Research that survives peer review

Every run captured, tagged, and replayable — including years from now.

Signed Local Traces

Every query stores prompt hash, model hash, parameters, and timestamp — exportable without revealing content.

Dataset Replay

Re-run a historical query set against any open-weight model. Keep comparisons honest and reproducible.

Version Pinning

Hold a matter or study on an exact model revision for the life of the engagement.

Security built for privileged material

Local inference, zero telemetry, signed audit logs, and air-gap support.

100% Local Inference
100% Local Inference
Zero Telemetry
Zero Telemetry
Air-Gap Ready
Air-Gap Ready
No-Training Guarantee
Access Controls
Access Controls
Signed Audit Logs
Signed Audit Logs

Frequently asked questions

Does any document content leave the device?
No. VirexaLLM runs the model locally. Prompts, documents, and responses never traverse the network. In air-gap mode, network access is blocked entirely.
How does this fit attorney-client privilege?
Because there's no third-party processor in the loop, privileged material stays within your firm's control. We never see, store, or transmit content.
Can we pin a model for a matter that takes years?
Yes. Pin to a specific GGUF model hash. Open-weight models don't get silently re-trained — your research six months from now will match your research today.
Is the audit trail defensible?
Every query and model load is recorded in a signed, tamper-evident local log with the model hash and quantization. Exportable for matter management and eDiscovery.
Can researchers reproduce experiments?
Yes. Because weights, quantization, and the runtime binary are all pinnable, any experiment is fully reproducible — critical for peer-reviewed work and expert testimony.

Your laptop is the server now

Download VirexaLLM and run Llama, Mistral, Phi-3, Gemma, or Qwen locally in minutes. Free desktop app for macOS, Windows, and Linux — your prompts never leave the device.