DeepSeek R1 Artifact and License Record

DeepSeek R1 Artifact and License Record

Family identity

DeepSeek's official R1 repository and paper distinguish several artifacts.

https://github.com/deepseek-ai/DeepSeek-R1

https://arxiv.org/abs/2501.12948

DeepSeek-R1-Zero used large-scale reinforcement learning without supervised fine-tuning as a preliminary step. The authors reported emergent reasoning behavior along with problems including repetition, poor readability, and language mixing.

DeepSeek-R1 added cold-start data and a multi-stage training path before and after reinforcement learning to improve reasoning and presentation.

Both large artifacts build from the DeepSeek-V3 base architecture. They should not be confused with the smaller distilled checkpoints.

Distilled artifacts

The repository lists six dense distilled checkpoints based on Qwen and Llama families. A distilled checkpoint is not the full R1 mixture-of-experts model and does not inherit every behavior, hardware requirement, or license condition from it.

The artifact name should preserve base family and size, such as a Qwen-based 32B distillation or Llama-based 70B distillation. Third-party quantization, conversion, runtime packaging, or fine-tuning creates another identity layer.

License chain

The repository states that its code and main model weights use the MIT License and support commercial use, modification, derivative works, and distillation.

https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE

The repository also states that distilled checkpoints retain obligations from their Qwen or Llama bases. A user must review the exact artifact, base license, code license, runtime, tokenizer, datasets or adapters, third-party package, and intended use.

"Open source," "open weight," and "commercial use" should not be applied to the company, hosted API, every dependency, or every derived package as if they were one license object.

Repository versus hosted service

Weights in a repository and a hosted API are different delivery models.

A local or private deployment controls the chosen artifact and runtime but creates hardware, security, privacy, patching, scaling, monitoring, and operations duties. A hosted API shifts some operations to the provider but adds current terms, service behavior, data policy, availability, rate limits, price, and provider dependency.

Current API source:

https://api-docs.deepseek.com/

Evaluation identity

Every evaluation should record checkpoint or API model name, source revision, date, license, quantization, tokenizer, runtime, hardware, prompt template, generation settings, tools, context, data, test set, scoring, repetitions, latency, tokens, cost, failures, and reviewers.

Author-reported benchmarks belong to the release record. They do not establish performance on another workload, language, risk class, quantization, runtime, or service.

Publication language

Use "DeepSeek-R1" only when the exact full artifact or paper-level family is intended. Name distilled checkpoints in full.

Use "open weight" when discussing available weights unless the complete artifact and license chain justify a stronger term.