For purchasing and procurement
One page with what a buyer needs: what it is, what it costs, what you get, how it is checked, and how to buy. Everything here can be verified from the public repository and its release page.
In one paragraph
OwnLLM is a private AI assistant that runs on your own computers and servers: nothing you type is sent to anyone. It also includes a deterministic engine that checks answers with proof gates where a gate exists. It is licensed per organization, not per user: unlimited people and devices, no usage metering, no license server, and it works offline. Organizations under US$5 million in annual revenue and all individuals use it free; larger organizations buy an annual license priced by revenue band.
Price list (annual, per organization)
| Your annual revenue (with affiliates) | Annual license | Includes |
|---|---|---|
| Under US$5 million | Free | Everything below |
| US$5 million to under US$25 million | US$1,900 | Everything below |
| US$25 million to under US$100 million | US$4,900 | Everything below |
| US$100 million to under US$500 million | US$9,900 | Everything below |
| US$500 million and above | US$19,900 | Everything below, plus a named contact for security and legal questions |
- No per-seat charge, no usage charge, no true-ups. Add people and computers freely.
- Evaluation is free for everyone, as long as you need: testing, development, learning and research never require a license.
- Cancel by not renewing. There is no lock-in: it runs offline, nothing phones home, and a license covers the versions you received (each version also becomes Apache-2.0 four years after release).
- Prices are in US dollars, for production use of OwnLLM's own code. Third-party parts keep their own (permissive) licenses and need no payment.
What you get
- Private by design. Listens only on the local computer; an API key protects the API; no telemetry; no auto-update; prompts never leave the machine. Downloads are checked against SHA-256 fingerprints.
- Runs on ordinary hardware. Four model sizes from 2 GB to 12 GB of memory, including a 7B model tuned for 8 GB laptops. Mac (Apple Silicon and Intel) and Linux (Intel/AMD and ARM).
- Deterministic engine. Proven answers first, rules next, a model only as a last resort, and a proof gate on every candidate; proven answers are remembered, re-checked and returned identically at no model cost.
- Your own notes, kept as compressed Markdown on disk and used only when a question needs them, with sources named.
- Structured answers that always follow a JSON Schema; repeatable answers (same question, same answer).
- Easy installation and removal. One command, no administrator rights, no other software; one command to remove.
- Source available, so your security team can read it.
How it is checked
| Evidence | Where |
|---|---|
| 101 automated tests of the installer and command line, and 59 of the deterministic engine | tests/, det/ (run on every change) |
| Exact bounds checker cross-checked against the Z3 solver on 1,500 random expressions | det/tests/z3_agreement.rs |
| Every release is installed and used on a clean Linux (Intel/AMD and ARM) and macOS machine before it can be published, and the 8 GB model is run in a bare container limited to 5 GB | .github/workflows/release.yml |
| Two independent security and licensing reviews before 1.1.0, with the findings fixed | the 1.1.0 and 1.1.1 history |
| Reproducible packages; pinned engine and model files, each with a SHA-256 | manifest/, release SHA256SUMS |
| Model comparison with the raw numbers (so you can repeat them) | docs/MODELS.md, docs/benchmarks/ |
| Published security policy and private vulnerability reporting | SECURITY.md |
Not claimed. OwnLLM has no third-party certification (such as SOC 2 or ISO 27001) and makes no compliance claim. Because data never leaves the computer, many questionnaire items about data handling are simple to answer: see below. Chat answers come from a small language model and can be wrong; the deterministic engine's guarantee applies to the problem types that have a gate (how it works).
Security questionnaire, short answers
| Question | Answer |
|---|---|
| Does the product send customer data to the vendor or any third party? | No. Prompts and usage are never sent. The only network use is downloading its own files, and an optional web search you turn on yourself. |
| Does it collect telemetry or analytics? | No. |
| Where is data stored? | On the computer that runs it (notes and proven answers as compressed files you can delete). |
| Does it phone home for licensing? | No. There are no license checks, activations or keys. |
| Does it need administrator rights? | No. It refuses to run as root. |
| Third-party components | llama.cpp (MIT), Qwen2.5 models (Apache-2.0), gzip-compressed files; list in THIRD_PARTY.md. |
| How are vulnerabilities reported and handled? | Privately, through GitHub; see SECURITY.md. |
| How are releases built and verified? | In public CI; packages are reproducible and fingerprinted; installers verify them before use. |
Cost compared with per-user AI assistants
Per-user assistant plans are commonly priced in the range of US$20 to US$30 per user per month; check current vendor prices. For illustration only, at US$25 per user per month:
| Users | Per-user plan, per year | OwnLLM, per year (for a company with US$60 million revenue) |
|---|---|---|
| 50 | US$15,000 | US$4,900 |
| 250 | US$75,000 | US$4,900 |
| 1,000 | US$300,000 | US$4,900 |
This compares price only. The largest cloud models are more capable than the models that run on your own hardware, so the saving is real when your work fits what OwnLLM does well. Evaluate it free first.
A note you can paste into a purchase request
We would like to license OwnLLM (https://github.com/ownllm-org/ownllm) for production use. It runs entirely on our own
hardware, sends no data to the vendor or third parties, and is licensed per organization rather than per user. The annual
fee for our revenue band is listed at https://ownllm.org/commercial/ and is a single fixed amount with no per-seat or
usage charges. We have evaluated it at no cost. The license text is at https://ownllm.org/licensing/.
How to buy
- Open an issue titled "Commercial license" in the repository, using the template (company name, revenue band; nothing private).
- We reply with a way to continue in private, a quote and an invoice for the annual fee in your band.
- You pay the invoice. The license applies from the date of payment to the versions you receive during the year.
Questions first? Ask in the same issue, or evaluate for free and ask later.