DeepSeek-V4-Pro-0813: MIT licence, weights included, no usage policy in the repository

In recent weeks we have reported the same trap more than once: a permissive licence on the card, and beside it a second document that empties it out — a usage policy ruling out defence and healthcare, a contract excluding the European territory. This time we ran the same check, file by file, and the trap isn’t there. It is worth writing up, because the buyer who has to decide doesn’t know that until they verify it themselves.

deepseek-ai/DeepSeek-V4-Pro-0813 appeared on Hugging Face on 13 August 2026. The API’s createdAt field, which we queried on 15 August, reads 2026-08-13T03:05:06Z; at that moment the repository showed 19,945 downloads, 486 likes and 92 files. The tags include deepseek_v4, text-generation, safetensors, fp8, 8-bit, license:mit and the reference arxiv:2606.19348.

The licence file, read in full

The repository’s LICENSE file opens with “MIT License Copyright (c) 2023 DeepSeek”, and continues with the grant: “Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software”. We downloaded it and checked it against the official MIT License text: it matches, word for word.

The weights are named, not just the code

A permissive licence on the code says nothing, on its own, about the weights: it is the point on which many inventories trip up. Here the README.md closes the question with a sentence that makes the check conclusive: “This repository and the model weights are licensed under the MIT License.” Not “the code”, not “the scripts”: the weights are named.

We then searched the README’s text for the phrases we have found elsewhere in recent days — usage policy, acceptable use, restrict, prohibit, shall not, must not, commercial use — and none appear. Among the repository’s 92 files, the only documents of the “read-me” kind are four: LICENSE, README.md, encoding/README.md and inference/README.md. No USAGE_POLICY.md, no restrictions attachment, no second file sitting beside the weights.

A record-keeping detail, not a thesis

One observation worth reporting for what it is, without stretching it: the licence file’s header reads “Copyright (c) 2023 DeepSeek” on a repository created in 2026. The licence file has evidently been reused from an earlier release. It doesn’t change the scope of what it grants — the text is still the full MIT License — but whoever archives this evidence should know the file carries a date different from the model’s own.

What the model says about itself

The README describes DeepSeek-V4-Pro-0813 as “the official release of DeepSeek-V4-Pro, superseding the preview version”, built “on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached”. It documents three reasoning_effort levels — low, high and max — and for the latter two recommends a maximum output of 384K tokens. The example command for serving the model with vLLM runs “on a single 4×GB300 node”: four accelerators on one node. For local execution it points to the inference/ folder; there is no Jinja chat template. The Technical Report is at arxiv.org/abs/2606.19348. We report none of the benchmarks that appear in its table: we have not verified them.

A policy, but elsewhere — and not on this

DeepSeek does publish terms of use for its own services: the “DeepSeek Terms of Use”, hosted on cdn.deepseek.com, define their scope as follows: “DeepSeek’s products and services include those provided to you through websites, applications […], software development kits (SDKs) […], application programming interfaces (APIs)”. The document never names Hugging Face repositories. It does name the weights, though, and this deserves stating precisely, because it is the kind of detail we usually find in others: clause 6.1, on intellectual property, says DeepSeek “holds all rights within the scope permitted by laws and regulations in this service (including but not limited to software, technology, programs, code, model weights and parameters, user interfaces, web pages, text, graphics, layout designs, trademarks, electronic documents, etc.)”. So the weights are there — inside a clause that bounds its own object to “this service”. We are not asserting those terms are irrelevant to someone downloading the model: that is a question for a lawyer, not for this article. We observe that the document defines itself around services DeepSeek provides, and that the weights published on Hugging Face carry an MIT licence naming them expressly.

See the service · Talk to an engineer

The problem we solve, and it’s the same one as always

The point isn’t that this particular licence is clean: it’s that no one can know that by looking at the label. You have to open LICENSE, check that it names the weights and not just the code, count the files, look for a second document under a different name. In recent weeks, on the same shelves, we found the exact opposite: an Apache 2.0 with a usage policy beside it ruling out defence and healthcare, a licence excluding the European Union, weights with no licence attached at all, a revenue threshold beyond which the open licence stops applying. Almost no organisation keeps a record of what it is actually allowed to do, today, with the models it has in production: whoever downloaded a model six months ago doesn’t know whether the licence has changed since, hasn’t kept the file in the version they downloaded, doesn’t know whether a policy sat beside the weights, and cannot prove any of it to an inspection or to a client in a tender.

What a clean licence doesn’t solve

A genuine MIT licence solves one problem — the right to use, modify and redistribute — and leaves others open, which we have not verified here and assert nothing about: where these weights run in production, on what hardware and in what numeric format; who answers for the provenance of the data the model was trained on; what constraints apply to a public body or to defence regardless of the software licence. These are questions distinct from the licence question, and should be kept distinct.

What we don’t know

We have not assessed the model’s quality and we report no benchmarks: we have not verified them. We don’t know whether the licence will stay this way: a licence is checked as of the date you download it, which is why it needs archiving alongside the weights, not recalled from memory months later. We have not checked export-control, public-procurement or national-security constraints that might apply to this model, and we assert neither their existence nor their absence. We don’t know whether DeepSeek’s terms of service bind, in some form, whoever redistributes output obtained from downloaded weights: that is not a question a repository can answer.

The two axes, applied to this case

Complying: the register of models in production becomes a check that runs on the client’s own systems — for each model, which version, downloaded when and from where, with which licence archived in the text of the day it was downloaded, which files sat beside it, who approved it and for what uses. With the dated record ready to show an inspector, a client in a tender, or a board.

Deciding: the same system brings models, data, contracts, archives and documents together into a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. A permissive licence on the weights like this one is what makes it possible to grow competence inside the client’s own perimeter — fine-tuning, adapting, integrating the model on the client’s own data — instead of pouring it into a third party’s model that the client doesn’t own, can’t inspect, and can’t take elsewhere. We always deliver in two modes: on-premise, on self-contained machines that require no deep integration into the client’s network, or dedicated cloud with a dedicated VPN and a data centre in Italy; always with shared management.

From the first session, at no cost, comes the dated register of the models you have in service: for each one, version, provenance, archived licence and approved uses — including the boxes that stay blank. It stays yours even if we don’t go on to work together.