Open-weight AI models in a clinic: what the Qwen, DeepSeek, Kimi and Llama licences say

The ScribeMD Team ·

Open-weight models let a clinic run capable AI on its own hardware. "Open" sounds like one licence, but it isn't. Qwen, DeepSeek, Kimi and Llama models come under different terms, and a model built on top of another one can carry two sets of terms at once.

For most clinics the conditions in these licences will never bite. You still want to know what they say before a model goes into your building, for the same reason you read any software agreement. This guide summarises the terms on the official Hugging Face pages for models a clinic is likely to run, as of when we checked them. Licences change, so read the current text before you rely on it, and involve your counsel for anything commercial.

The quick reference

  • Qwen3-30B-A3B (Alibaba's Qwen team): Apache 2.0 (model page).
  • Qwen2.5-72B-Instruct: the Qwen licence, not Apache. It requires anyone commercially using the model in a product or service with more than 100 million monthly active users to request a licence (model page).
  • DeepSeek-R1: MIT (model page).
  • DeepSeek-R1-Distill-Llama-70B: listed as MIT, but derived from Llama 3.3 70B Instruct, which is "originally licensed under llama3.3 license" (model page).
  • Kimi K2 (Moonshot AI): a modified MIT licence with an attribution condition at large scale (model page).
  • Kimi-Dev-72B: listed as MIT, with Qwen2.5-72B as its base model (model page).

Lesson one: check the base model

Many of the models a clinic can actually fit on one machine are fine-tunes or distillations of other models. Their own licence covers the work their authors added. The model they started from has a licence too.

DeepSeek is unusually clear about this. Its R1 model page says the distilled Qwen versions "are derived from Qwen-2.5 series, which are originally licensed under Apache 2.0 License," and that DeepSeek-R1-Distill-Llama-70B "is derived from Llama3.3-70B-Instruct and is originally licensed under llama3.3 license." So a model page that says MIT at the top can still point you to a second licence further down.

Kimi-Dev-72B shows the same pattern from a different direction. Its page lists MIT, and its model tree names Qwen2.5-72B as the base model. Qwen2.5-72B's own page lists the Qwen licence. Read both.

Lesson two: most scale conditions won't reach a clinic

The conditions that trouble large tech companies are written for products with enormous user numbers.

The Qwen licence on Qwen2.5-72B-Instruct asks commercial users whose product or service has "more than 100 million monthly active users" to request a licence. Kimi K2's modified MIT licence adds one condition to standard MIT: if the software is used in commercial products or services with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, you have to display "Kimi K2" prominently on the product's user interface (Kimi K2 licence).

A clinic using a model internally is nowhere near those thresholds. A software company building a product on these models for thousands of clinics should still read them closely, and so should a large health system.

Lesson three: the flagship and the box are different models

When people say "DeepSeek" or "Kimi," they often mean the flagship. Those are very large. On Hugging Face, DeepSeek-R1's weights total about 690 GB, and Kimi K2's total about 1 TB. A clinic server with 128 GB of memory can't hold either.

What a clinic runs is a smaller member of the family: a 30B mixture-of-experts Qwen model, a 70B DeepSeek distillation, a 72B Kimi variant. Those are different models with different licence lines, as the list above shows. When a vendor says a box "runs DeepSeek," ask which model and which version.

On Cortex, with 128 GB of unified memory, NVIDIA's guidance for the GB10 chip is inference on models up to 200 billion parameters (NVIDIA, DGX Spark). What fits in practice depends on the model's size, quantization and context length, and on leaving room for the ScribeMD scribe running alongside it.

Lesson four: a licence isn't a clinical sign-off

An Apache or MIT licence tells you what you may do with the weights. It says nothing about whether the model is accurate enough for a clinical task. None of these licences is a statement about clinical accuracy.

Use them the way you would use any general tool in a clinic. Keep a person responsible for any output that reaches a patient or a chart, and don't let a general model's answer stand in for clinical judgement. In ScribeMD, notes written on Cortex wait for the clinician to review and sign. The same principle should apply to anything else you run on the box.

Keep a model register

One simple habit covers most of this. For each model on your server, record:

  1. The exact model name and version, as it appears on the source page.
  2. Where you downloaded it, and a checksum of the files.
  3. Its licence, and the licence of its base model if it has one.
  4. The date it was loaded, and who approved it.
  5. What it is used for in the clinic, and what it isn't used for.
  6. Who reviews its output before it reaches a patient or the chart.

Because a self-hosted model changes only when you change it, the register stays accurate until you decide otherwise. That's harder to promise with a cloud model that a vendor can update overnight.

A note on names

Qwen, DeepSeek, Kimi and Llama are trademarks of their owners. Those companies aren't affiliated with ScribeMD and don't endorse Cortex. We name their models because clinics ask about them.

Next steps

If you are weighing a server for your clinic, our on-premise checklist covers placement, maintenance and fallback. For a single clinician who wants the scribe on their own laptop instead, ScribeMD Desktop runs a smaller model locally and is free for individual doctors.