Who this is for
Law firms, accounting and wealth management firms, insurance brokers, healthcare administrators, and any company with confidentiality obligations to its clients or a contract that forbids third-party processing. Also companies that simply do not want their operating data used by anyone else.
What private means here
The model runs on a server in your office, in your data center or in a cloud account that only you administer. Prompts, documents and outputs stay inside it. Nothing is sent to OpenAI, Anthropic, Google or any other AI provider. You hold the keys. We document the boundary so your auditor or your client can read it.
What we deploy
Current open-weight models from labs such as Alibaba (Qwen), Moonshot (Kimi) and others, chosen for the job: reading documents, drafting, classifying, searching your own files. We benchmark candidates on your actual work before choosing. Fine-tuning and retrieval over your documents when the task justifies it, not by default.
Hardware, software and hosting
- Sizing. How many people, which workloads, what response time. Then a hardware spec you can read.
- Procurement. We can buy and configure the hardware, or hand the spec to your IT team and they buy it. Either way it is yours.
- Hosting. On your premises, in your colocation, in a private cloud account you own, or on infrastructure we run for you under a separate agreement. We recommend based on your team and your obligations.
- Deployment. Model serving, access tied to your existing logins, logging, backups and an upgrade path.
- Support. Monitoring, model updates and the same managed service that covers our other builds.
What it costs
Scoped after discovery. Hardware is quoted separately, at cost through us or through your vendor. We show the running cost next to what the same workload would cost at a cloud provider, so the decision is a number, not a feeling.
Where the limits are
Open-weight models are very good and a step behind the frontier on the hardest reasoning tasks. For most business workflows the difference does not matter. For some it does. We test yours before you commit, and we will say so if the private option is the wrong fit. We do not make HIPAA, SOC 2 or other certification claims. We build to the requirements your obligations set and document what was done.
Related: AI strategy and roadmap and managed AI agents. See also how we handle your data.
Can we mix private and cloud models?
Yes. Many firms route sensitive work to the private model and everything else to a cloud model. We set the routing rule with you and document it.
Can you host it for us?
Yes, under a separate hosting agreement, if your obligations allow infrastructure we run. Many clients prefer their own premises or cloud account, and we support that too.
Will our staff notice a difference?
Usually not. The work shows up in the same tools. Response speed depends on the hardware, which is part of the sizing.
How do model updates work?
Newer open-weight models arrive often. We test the candidate on your work, show you the difference, and update on your approval.
