On 20 January 2025 DeepSeek released the weights of DeepSeek-R1 under the MIT licence, alongside a technical report and six distilled models. For a reasoning system put up against OpenAI’s o1 on maths and coding benchmarks, a licence this permissive on the parameters is rare. It is worth pinning down exactly what that licence covers and what it does not: the difference matters to anyone who has to verify the model before taking it into production.

Context

DeepSeek is a Chinese research company founded in 2023 as an offshoot of the quantitative fund High-Flyer. In December 2024 it published DeepSeek-V3, a Mixture-of-Experts model with 671 billion total parameters, 37 billion activated per token, trained — according to the technical report (arXiv 2412.19437) — in 2.788 million GPU-hours on H800 accelerators. DeepSeek-R1 arrived on 20 January, built on top of V3-Base and specialised for reasoning.

What drew attention was the manner of the release, more than the reported performance: downloadable weights, a technical report that goes into the detail of the training procedure, and — for R1 — the MIT licence extended to the model weights as well. That happens less often than the open source label suggests.

Architecture

R1 comes out of a multi-stage pipeline described in the report. The starting point is DeepSeek-R1-Zero: the base model is trained with pure reinforcement learning (RL), with no preliminary supervised fine-tuning (SFT) step, using a reward function anchored to verifiable criteria — answer correctness in maths, tests passed for code. R1-Zero develops explicit reasoning chains, but it has readability problems and mixes languages.

R1 fixes these flaws: it introduces cold-start data before RL and alternates two RL stages with two SFT stages. The model lays out an intermediate chain of thought — it breaks the problem down, checks its consistency — before the final answer. Six distillations, from 1.5 to 70 billion parameters on Qwen and Llama bases, carry part of these capabilities over to models that run on modest hardware.

Underneath, the V3 architecture uses Multi-head Latent Attention (MLA) to shrink the key-value cache and an auxiliary-loss-free strategy for expert load balancing. The report documents them and makes them reproducible at the conceptual level: they do not stop at a benchmark number.

The critical point

The licence is not uniform across the release. For R1 the MIT terms hold for both the repository code and the model weights: commercial use, modification and distillation stay free, with no use restrictions. For V3, by contrast, the repository code is MIT, but the weights fall under a separate model licence derived from OpenRAIL, with responsible-use clauses. Two releases from the same company, weeks apart, with different licensing regimes on the parameters.

There is then a second layer, which has nothing to do with the licence. Open weights does not mean reproducible training. What DeepSeek makes available is the final parameters and a report describing the procedure. Left out: the pre-training dataset (14.8 trillion tokens, described but not distributed), the exact code of the RL pipeline, the make-up of the cold-start data, the internal evaluation sets. A third party cannot re-run the training and get the same model back; it can only take the weights as they are and trust the report for everything else.

This is where open weights and open source part ways, in the sense open source carries for software. In software, open source means access to the preferred form for making modifications — the source. For a model, that form would include the training data and code; the weights on their own look more like a compiled binary than source. The Open Source Initiative published version 1.0 of the Open Source AI Definition in October 2024 precisely to put this distinction in order, and it asks for sufficient information about the data from anyone claiming the label. R1’s MIT weights, on their own, are not enough to meet that requirement.

Implications

For anyone deciding whether to adopt R1, the distinction has practical consequences. The MIT licence on the weights closes the legal question of commercial use and redistribution: one constraint fewer. It does not close the question of auditability. Without the dataset and the pipeline you cannot independently verify what the model saw in training, nor reproduce the stated properties from scratch; only downstream checks on the available weights remain open — red-teaming, evals on your own sets, behaviour analysis.

For those working under regulatory constraints this is the point. The EU AI Act, in force since 1 August 2024, requires general-purpose models to keep technical documentation and — under Article 53(1)(d) — to publish a sufficiently detailed summary of the content used for training, following a template provided by the AI Office. An open-weights release under a permissive licence helps, but it does not substitute for that documentation: the transparency required concerns the data, and an MIT licence on the parameters does not make that data available.

There is also a reading about the economics of the field. The training cost reported for V3 — roughly 5.576 million dollars, computed at 2 dollars per H800 GPU-hour over the stated 2.788 million hours — is a fraction of current estimates for Western frontier models. The figure covers the final training run only: no research, no failed experiments, no hardware. The point that remains is a different one: architectural and algorithmic efficiency can lower the compute barrier, and an actor with limited resources relative to the large labs can still release a competitive model with downloadable weights.

Limits

Almost everything said about R1 at the time of release rests on the technical report and on benchmarks declared by the author. Independent performance verification takes time and uncontaminated evaluation sets; the cost figures do not compare directly with those of competitors, who use different accounting methods that are rarely public. The MIT licence on R1’s weights can be verified by reading the file in the repository; the make-up of the training cannot, and on that one relies on what is written. Keeping the two apart is the minimum, before treating open weights as a synonym for transparency.

Cover image: DeepSeek logo: a stylised whale icon next to the DeepSeek wordmark, rendered in purple on a light background — logo by RoadMaster19, CC BY 4.0 — https://commons.wikimedia.org/wiki/File:DeepSeek_purple.png