DeepSeek open weights and distilled students for adaptation

Who this page is for

Read this page when the brief already says DeepSeek and someone has to write down which repository, which grant, and whether you intend to serve a teacher, adapt a published student, or distill later. Typical readers are counsel who were told “DeepSeek is MIT”, a platform owner who just watched a 671B-class demo, and a specialist who will judge extracts or attributes. The page is not a ranking of DeepSeek against Llama or Qwen. It is not a reprint of blog cost-per-token tables. It is not a tutorial for running a 671B mixture-of-experts.

If you have not yet chosen a family, start at the models catalogue and the openness comparison. If the constraint is Meta’s community terms after adaptation, go to Llama — and do not assume R1’s distillation grant exists there. If the constraint is an Apache file on a commercially common dense card, go to Qwen or the current Mistral 3 announcement. If the constraint is a reproduced training stack, go to OLMo. If the constraint is the method of turning a teacher into a student, go to the distillation page after this one. DeepSeek’s job on this site is the R1 MIT-plus-distillation grant versus the V3 Model License versus the mistake of serving the teacher by default.

A useful reader can already say teacher, published student, or later own-student in one sentence. If that sentence is still “put DeepSeek on our policies”, you do not yet have a project. You have a demo. We will not invent the identifier from the family name.

R1 MIT is not the V3 Model License, and neither is a serving plan

If the brief says “DeepSeek is MIT”, ask which repository. The cited R1 source states the code repository and the model weights are MIT-licensed, support commercial use, and allow modifications and derivative works including distillation for training other LLMs. That is a real grant. It is the reason this family appears in distillation conversations at all. It is not a grant you may paste onto V3, and it is not a grant you may paste onto a student file you have not opened.

The cited V3 source states the code repository is MIT-licensed while use of the V3 Base/Chat models is subject to the Model License. The series supports commercial use. That is still an open-weight story you can often run and adapt, subject to LICENSE-MODEL. It is not the same sentence as “the weights are MIT”. Counsel who approved R1 last month have not automatically approved V3. A later adapter does not convert LICENSE-MODEL into MIT.

Both published tables list 671B total parameters, 37B activated, and 128K context. That is a hardware fact. It is not a quality result on your workflow. A team that can serve a 14B dense card may not be able to serve a 671B-class mixture-of-experts. Write the serving plan before the training plan. Do not treat the context window as evidence the model will read your binder. Do not treat 37B activated as “really a 37B model” for procurement slides. The box still has to host the MoE.

The R1 distillation grant is permission, not a project

R1’s published MIT line is unusual among the families on this site: distillation into other LLMs is explicitly permitted on that repository. That is why people arrive here from a distillation brief. Permission is not a schedule. Permission is not a quality bar on the teacher. Permission is not a student card. You still need a licensed teacher that already meets the written bar on unused test documents, permission to use the traces you will train on, and a read of the family you intend to pour those traces into.

If the intended student is another DeepSeek file, open that file. Published students inherit a reasoning style from a larger teacher in the vendor’s write-up. They do not inherit R1 MIT by family name. They do not inherit your book. Using one as a base is evaluation or fine-tuning, not “we distilled our policies”. If the intended student is Llama, stop and read /models/llama before anyone records traces. R1’s grant does not rewrite another family’s agreement.

Own distillation is a later engagement. The honest sequence is: name the teacher, name the grant, measure the teacher on unused test documents, then talk about a student. Skipping to traces because the repository said distillation is allowed is how you spend a quarter training a student of a teacher that never met the bar. The distillation page owns the method. This page owns which DeepSeek grant you thought travelled with the traces.

V3 is commercial, not MIT on the model

V3’s code being MIT is a software fact about the repository. The model sitting under LICENSE-MODEL is the fact that binds the weights. The series supports commercial use. That sentence is useful. It is not a substitute for counsel reading LICENSE-MODEL for redistribution, acceptable use and whatever else that file actually says. We will not paraphrase a Model License into MIT so a slide stays tidy.

A team that wants “the DeepSeek 671B” without saying R1 or V3 is asking counsel to sign two files. Write the repository URL into the statement of work. If they later swap R1 for V3 because a serving stack was easier, that swap is a new legal read. Size did not stay the same in a way that matters to the grant. The table stayed the same. The licence line did not.

OSI’s Open Source AI Definition still asks for the preferred form to modify the system. OSI’s open-weights note still says parameters are not that form. Downloadable DeepSeek weights, MIT or Model License, let you host and adapt a model. They do not, from the cited repositories alone, give you the training data and recipe that definition treats as the preferred form. If procurement meant OSAID, DeepSeek is the wrong exhibit. Point them at OLMo.

Size bands change the box, not the repository split

The teacher band is the published 671B total / 37B activated MoE with 128K context. That band is an inference plan before it is an adaptation plan. The student band is smaller cards published by DeepSeek or by other people who used a DeepSeek teacher. Those cards must be read individually. Do not inherit R1 MIT by family name. Do not inherit V3’s commercial sentence by family name. Do not inherit a quality result from a public reasoning demo.

Pick the band from hardware you actually operate and from a written quality bar. Do not pick the teacher because a listicle said it won maths. Do not pick a student because it is cheaper on a spreadsheet that assumed it would meet the bar. A student that almost meets the bar is not a distillation success. It is unfinished work. Whether you should shrink at all still lives on the commercial /slm page. This page is where “which DeepSeek grant, and are we serving the teacher?” lives.

A published student is someone else’s student

DeepSeek has published students that inherit reasoning-style behaviour from a larger teacher. That sentence is a vendor fact. It is not your specialist model. It is not “we distilled our book”. Your insurance extract or catalogue attribute job still needs examples, unused test documents, and a comparison with the current process. Thinking traces on public maths exams are not source-linked policy fields. They are not SKU attributes.

Using a published student as a base can be the right first commercial move: you evaluate or fine-tune a smaller file you can actually serve, under whatever licence that file states. Write the identifier. Write the licence line from that file. Do not tell a steering group you distilled. You downloaded. Distillation is the later engagement, after the teacher you are licensed to use already meets the bar and counsel has read both grants.

Students trained by other organisations on DeepSeek traces are a fourth pile. They are not R1. They are not your unused pack. They are someone else’s artefact with someone else’s data story. If a vendor offers “DeepSeek-distilled for insurance”, ask whose teacher, whose grant, whose examples, and whose unused documents. We will not treat a marketplace label as a result.

What changes in an adaptation run

On a named R1 or V3 teacher, a typical specialist run is usually the wrong first run. The published tables are 671B-class MoE. Unless the organisation already operates that class, the honest first artefact is not an adapter on the teacher. It is a decision: evaluate a published student, pick another family’s smaller card, or plan distillation as a later engagement. If you do freeze a teacher you can actually serve, the artefacts are the adapter or merge, the exact identifier, the MIT text or LICENSE-MODEL text that applied, and notes from unused test documents. The grant does not change because the optimiser ran.

On a published student, the run looks like any other specialist adaptation: freeze the named smaller card, train adapters, store the identifier, that card’s licence line, and unused-test notes. You have not performed distillation. You have fine-tuned someone else’s student. Say that in the handover. If the team later wants their own student, that is a new statement of work: teacher grant, teacher quality bar, trace permission, student family. The LoRA page owns the adapter file. Do not ask this page for QLoRA mechanics. This page owns which DeepSeek grant you thought travelled with the file.

Inference location does not change because adapters exist. A hosted DeepSeek-shaped SKU is still someone else’s building. Open weights only help when the GPU environment is named. 128K context on the table does not move the box into your region.

Insurance: do not serve the teacher by default on policies

Retail: a student is an attribute base only after unused tests

When DeepSeek is the wrong first conversation

Pick another family when the organisation will not read more than one file and the brief only says DeepSeek. Pick another family when counsel asked for OSAID and will not accept open weights plus MIT or a Model License. Pick another family when the only available path is a hosted 671B-shaped SKU and the requirement was take-home adapters on a box you operate. Those are identifier and inference problems. They are not solved by a better expert-routing config.

DeepSeek is also the wrong first conversation when someone wants to serve the teacher by default on policies or catalogue traffic they have not measured. It is the wrong first conversation when they want to call a downloaded student “our distillation”. It is the wrong first conversation when they intend to pour traces into a family whose agreement forbids that use. Read the other family’s page. R1’s grant does not delete it.

What handover must include

A useful DeepSeek handover states teacher, published student, or later own-student in the first line. Then it includes the repository or card URL, the MIT text or LICENSE-MODEL text that actually applied, adapters only if weights were actually trained, and evaluation notes from unused test documents. It is not a claim that “DeepSeek is MIT”, that a published student is your distillation, that inference is private, or that a public reasoning score transferred to your book.

Operators still need the base weights and a path that allows the stack. Serving 671B total versus a smaller student changes the operator’s week. Merging simplifies a single-task image. Keeping adapters separate helps more than one product on one frozen student. Choose from the operating plan. If the track was a hosted SKU, the handover is credentials and an exit sentence — not a weight file you thought was R1 MIT.

Procurement questions that belong in the brief

Is this R1, R1-Zero, V3, a published student, or a hosted SKU? If R1, have we kept the MIT-plus-distillation sentence and the 671B/37B table? If V3, have we kept LICENSE-MODEL as the model grant, not the code grant? If a student, have we opened that file instead of inheriting R1? Will we serve the teacher, or is that the wrong default? Did anyone mean OSAID when they said open source? Have unused test documents been reserved? Is own distillation in this statement of work or a later one?

Those questions are more useful than a debate about whether a reasoning demo “beats” another lab. This page exists so the R1 grant, the V3 Model License and the teacher-versus-student serving decision do not collapse into one word on a slide.

Sources

1. DeepSeek, DeepSeek-R1 repository (Accessed September 2026). DeepSeek-R1 states the code repository and model weights are MIT-licensed, support commercial use, and allow modifications and derivative works including distillation for training other LLMs. R1 and R1-Zero are listed as 671B total / 37B activated with 128K context.
2. DeepSeek, DeepSeek-V3 repository (Accessed September 2026). DeepSeek-V3 states the code repository is MIT-licensed while use of the V3 Base/Chat models is subject to the Model License. The published table lists 671B total / 37B activated parameters and 128K context. The series supports commercial use.
3. Open Source Initiative, Open Source AI Definition 1.0 (Accessed September 2026). An open-source AI system must provide the preferred form for modification, including parameters under OSI-approved terms plus the data information and code needed to derive them.
4. Open Source Initiative, Open Weights (Accessed September 2026). Open weights are the published parameters of a trained network. Weights alone are not the preferred form for modifying the system.

Questions we hear first