Comparison Reviewed September 2026

Open weights or proprietary?

Two years ago this was a capability question with an obvious answer. It is now a deployment and licensing question with a genuine answer on both sides, and the capability gap has narrowed to the point where it decides fewer cases than anything else on this page.

SIDE BY SIDE

	Open weights	Proprietary API
Run on your own hardware	✓	—
Absolute frontier capability	—	✓
Pin a version indefinitely	✓	—
Fine-tune and keep the artefact	✓	Platform-bound
Cost model	Fixed infrastructure	Per token
Zero-effort scaling	—	✓
Licence complexity	Varies — read it	Terms of service
Adds a sub-processor	—	✓

The structural differences. Note that most rows are not about model quality at all, which is the point.

THE LICENCE IS NOT A FORMALITY

“Open weights” covers a wide range of terms and they are not interchangeable. Some models ship under Apache-2.0, which is a standard permissive licence your legal team already recognises. Others carry bespoke community licences with attribution requirements, user-count thresholds or restrictions on training other models. A few are research-only and cannot be used commercially at all.

Check the licence for the specific model and version you intend to deploy, record which one you chose, and keep that record. “Which licence did we deploy under” is a question that gets asked eighteen months later by someone doing due diligence, and reconstructing the answer is far harder than writing it down at the time.

Note too that a model with published weights is not necessarily open source by the OSI definition. The distinction matters where procurement policy is written around recognised licences, and it is worth being precise rather than using the terms interchangeably.

THE COST INVERSION

At low volume, proprietary APIs are cheaper — there is no hardware and no operator. At high sustained volume that inverts, because one side scales linearly with usage and the other does not. Where the crossover sits depends on your volume, your model size and your engineering costs, and it is arithmetic rather than a matter of allegiance.

What is frequently missed is the second-order effect. Once marginal cost approaches zero, the set of things worth processing expands. Work that could not justify a per-token charge — reviewing every document rather than a sample, embedding an entire archive — becomes viable. In our experience that expansion, rather than the saving on existing usage, is where the larger value sits.

QUESTIONS

QUESTIONS — 3

Narrow for the majority of production tasks and still real at the hardest end — complex multi-step agentic work and the most difficult reasoning. The practical consequence is not choosing a side but routing: open weights for the bulk, frontier for the minority that needs it.

Llama for ecosystem depth, Mistral for EU domicile and Apache-2.0 terms on several of its models, DeepSeek for reasoning economics, Qwen for size range and multilingual coverage. We benchmark two or three against your task rather than recommending from reputation.

Different, not less. You take on patching and infrastructure security; you remove an entire third party from the data path. For most regulated organisations the second outweighs the first, which is why we would expect self-hosting to concentrate at the regulated end of the market.

Keep reading

30-minute technical call · no deck