DeepSeek R1: open weights do not remove operating responsibility

DeepSeek announced R1 on 20 January 2025; the original research paper was submitted on 22 January. The paper distinguished R1-Zero, trained with reinforcement learning without preliminary supervised fine-tuning, from R1, which added cold-start data and a multistage training process. Its comparisons with OpenAI o1 were developer-reported benchmark findings, not universal equivalence across business tasks. The release is historically important because downloadable reasoning models widened deployment choices. It does not establish a current ranking or current service terms. 1 2

The original paper described six dense distilled models based on Qwen and Llama, ranging from 1.5 billion to 70 billion parameters. These are not simply smaller copies with guaranteed identical behaviour. Select the exact checkpoint before comparing memory requirements, quality or license terms. A successful trial of one distilled model does not validate the full R1 system, and the reverse also holds. Record base-model dependencies and preserve the tested artifact so a later download cannot silently change the decision.

Downloading weights, using a third-party hosted endpoint and using the developer’s public chat service create different data paths. Do not transfer a privacy conclusion from one to another. Self-hosting gives infrastructure control but also assigns responsibility for access management, logging, updates and incident response. External hosting requires checking retention, subprocessors and transfer arrangements. For an enterprise, the meaningful question is where a particular prompt and its retrieved documents travel, not whether the model family is casually described as open source.

The January release presented R1 code and weights under the MIT license. That statement should not replace review of the exact distributed package, including notices inherited from distilled base models. Open weights do not demonstrate that all training data is disclosed, and a permissive software license is not a warranty about every possible output. Connect the review to provider obligations and AI security, keeping intellectual-property questions separate from runtime security and data protection.

The original paper noted readability, repetition and language-mixing problems in R1-Zero. That makes it especially important not to treat emergent reasoning as automatic operational readiness. Test Swedish and English task pairs, incomplete questions and deliberately contradictory inputs. Measure cost per accepted answer, including reasoning-token generation and retries. A model that knows when to request clarification can be more useful than one that produces a fluent but unsupported solution. Preserve failure examples rather than reporting only an average score.

Use agent permissions to define tools and actions independently of model capability. A pilot should have a rollback owner, an approved document set and explicit escalation conditions. Open reasoning is a deployment option, not an exemption from governance. This article is published on 25 October 2026 as a retrospective of the January 2025 release and original paper. It does not verify current API prices, endpoint availability, newer R1 checkpoints or present-day benchmark leadership. Those require separately dated checks before procurement.