This comparison is worth making specifically, because DeepSeek is the model most often proposed as the moment self-hosting became credible — and because it carries a governance question the other open families raise less sharply.
Our updated DeepSeek V4.1-Flash guide covers the September release, image support, open weights and a dated Flash-versus-Pro API pricing comparison. DeepSeek also offers a hosted API; this page specifically compares operating the weights yourself with buying a hosted model service.
| Criterion | DeepSeek, self-hosted | Frontier API |
|---|---|---|
| Data leaves your infrastructure | No | Yes |
| Cost per request at volume | Hardware, energy and operations | Per token, linear |
| Hardest reasoning and agentic work | Evaluate the pinned release | Evaluate the pinned release |
| Version stability | You control it | Provider controls it |
| Fine-tuning and distillation | Permitted, artefact is yours | Platform-bound if offered |
| Setup effort | Weeks | Hours |
| Provider jurisdiction relevant | Supply-chain review still applies | Yes |
| Governance review needed | Model provenance | Data transfer |
The mixture-of-experts architecture gives inference economics that make self-hosting viable at a scale where dense models of comparable quality would not be — throughput per unit of compute is high even though memory requirements remain substantial. That is the technical fact behind the commercial argument.
DeepSeek’s recent releases have shipped under the MIT licence at the time of writing, permitting commercial use, modification and distillation — which means the path from “a large model that solves our task” to “a small model that solves our task at a tenth of the cost” is lawful and open. Licences are set per version, so check the licence file of the exact release you deploy. For high-volume narrow work, that path is frequently the entire project.
And you gain version permanence. Nobody deprecates a file on your server. For any system that went through formal validation, that is worth more than a benchmark point.
The absolute ceiling on the hardest work. On complex multi-step agentic tasks the leading proprietary models retain an edge, and it is visible if that is your workload. For most production tasks it is not, which is testable in a day rather than arguable.
Operational simplicity. An API is a URL; a self-hosted deployment is infrastructure with an owner, a patch cycle and an on-call story. Pricing that honestly is what turns a good decision into a good experience.
And you take on a governance review rather than avoiding one. Self-hosting closes the data-transfer question completely, and it opens a model-provenance question: who trained this, what does the licence permit, how does it behave on sensitive topics, what did we test. That review is straightforward and it is not nothing — the model supply chain page sets out how we structure it.
Self-hosted, the data-transfer objection does not apply — the weights are a file and transmit nothing. What remains is behavioural testing on your own material and a documented provenance assessment, both of which are routine. Where an organisation has a standing policy against models of Chinese origin, that policy is the answer and Mistral or Llama occupy adjacent ground.
Technically yes and for most European enterprise clients we would not recommend it without a completed transfer assessment, because that reintroduces exactly the jurisdiction question self-hosting removes. The interesting case for this model is on your own hardware.
Benchmark both on your task with your data, and price both at your projected volume including operator time. It is a few days of work and it converts an argument into a decision. We do this as a first engagement regularly, and it sometimes concludes against self-hosting.
A few days against your real task settles this permanently, in whichever direction the numbers point.
30-minute technical call · no deck
What we built, what stage it is at, and a link to the live thing wherever there is one.
yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.
Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.
Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.
Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.
Apps, platforms and campaigns for these teams, long before the AI work. Every one is written up on the work page — what we did, and where our part ended.
