The Short Answer
Most foundation model vendors will not let you run their model on your own hardware — Anthropic, OpenAI's frontier GPT models, and Google Gemini are API-only — while the ones that will split into open-weight families you download and self-host and Cohere, which sells contracted private deployment, and on ibl.ai you own all the code and the data around whichever you choose.
The open-weight tier is crowded and genuinely good: Meta, Mistral, Alibaba, DeepSeek, Google, Microsoft, IBM, NVIDIA, and OpenAI all publish weights you can run air-gapped.
The decision buyers get wrong is treating this as one choice. A model is not a platform. ibl.ai is model-agnostic across every vendor below, carries no per-seat pricing, and deploys anywhere — your cloud, on-premise, GovCloud, or fully air-gapped.
Which foundation model vendors can be deployed on-premise?
Three tiers, and the distinction between them is the whole decision:
| Vendor | Model family | On-premise? | License shape |
|---|---|---|---|
| Meta | Llama | Yes — open weights | Llama Community License (MAU threshold applies) |
| Mistral AI | Mistral / Mixtral | Yes — open weights + commercial on-prem | Apache 2.0 on several; commercial for others |
| Alibaba | Qwen | Yes — open weights | Apache 2.0 on much of the family |
| DeepSeek | DeepSeek / R1 | Yes — open weights | MIT on R1 |
| IBM | Granite | Yes — open weights | Apache 2.0 |
| Microsoft | Phi | Yes — open weights | MIT |
| NVIDIA | Nemotron + NIM | Yes — open models, on-prem inference | Open model license; NIM under enterprise terms |
| Gemma (not Gemini) | Gemma yes; Gemini no | Gemma Terms of Use | |
| OpenAI | gpt-oss (not frontier GPT) | gpt-oss yes; GPT no | Apache 2.0 on gpt-oss |
| Cohere | Command | Yes — contracted private deployment | Commercial; managed access, not ownership |
| Anthropic | Claude | No — API only | Commercial API terms |
| ibl.ai | Platform — runs any of the above | Yes — incl. GovCloud and air-gapped | Perpetual license; you own all the code and the data |
Model lineups and license terms change frequently. Verify the current terms with the vendor before you contract.
What is the difference between open weights and open source?
Open weights means the vendor publishes the trained parameters so you can download and run the model yourself. It does not mean the model is open source in the OSI sense.
The clearest example is Meta's Llama Community License, which is not an OSI-approved open-source license. It carries an acceptable-use policy and a condition requiring a separate license for products exceeding 700 million monthly active users.
Apache 2.0 and MIT — which cover Qwen, Granite, Phi, DeepSeek's R1, and OpenAI's gpt-oss — are genuinely permissive. If your procurement team needs an OSI-approved license, that is the tier to look at.
None of these licenses say anything about the software around the model. That is a separate question, and it is the one that determines what you actually own.
Does deploying a model on-premise mean you own your AI stack?
No, and this is the most expensive misunderstanding in the category.
Downloading open weights gets you a model that turns tokens into tokens. It does not get you agents that take actions, retrieval over your documents, authentication, role-based access control, audit logging, PII redaction, guardrails, or a support contract.
That application layer is where the engineering cost actually lands. Teams that budget for "we'll just run Llama on-prem" routinely discover they have committed to building a platform.
So there are two ownership questions, not one. Who owns the model weights, and who owns the code and data of the system around them? A vendor-hosted application layer re-creates the exact dependency the on-premise model was meant to remove.
Where does ibl.ai fit among on-premise foundation model vendors?
ibl.ai is not a foundation model vendor — it is the platform layer that runs any of them on infrastructure you own. That is a deliberate position, because model leadership changes every few months and an application stack welded to one vendor inherits that vendor's roadmap.
On ibl.ai you own all the code and the data. Full source code under a perpetual license on your infrastructure — not API access to someone else's platform.
It is model-agnostic across every vendor in the table above, including Cohere's own Command, so you route each workload to whatever model leads today and switch without rewriting the platform.
There is no per-seat pricing. Billing is usage-based against a budget cap you set, so cost tracks what your organization actually uses rather than how many people you employ.
And it deploys anywhere: your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
How should a regulated buyer choose between these vendors?
Start by separating the two decisions. Pick the model on capability, latency, and cost for your workload. Pick the platform on ownership, deployment reach, and whether it locks you to one model.
Then apply your compliance constraint. If your deployment must be air-gapped, the contracted-private tier still involves a vendor relationship at runtime, while open weights on hardware you control do not.
Ask every vendor the same three questions. Can this run with no outbound network connectivity? Do we receive source code, or access? If we switch the underlying model in eighteen months, what breaks?
Run a proof of concept on your actual data, in your actual network posture. Air-gapped behavior is the thing demos are least likely to represent honestly.
Explore the platform on the Agentic OS page, the enterprise solutions overview, or the self-hosted enterprise AI platform comparison.