What DeepSeek R1 Is: Models, Weights, and API

What DeepSeek R1 Is

DeepSeek R1 is a reasoning-model family introduced by DeepSeek in January 2025. The name covers a full R1 model, an earlier R1-Zero research artifact, and smaller dense checkpoints distilled into Qwen and Llama base families.

It should not be used as if those artifacts, third-party quantizations, and the hosted DeepSeek API were one model with one license, one hardware profile, and one behavior.

R1-Zero came first

DeepSeek's R1 paper describes DeepSeek-R1-Zero as a model trained through large-scale reinforcement learning without supervised fine-tuning as a preliminary step.

The authors reported emergent reasoning behavior. They also reported problems such as repetition, poor readability, and language mixing.

R1-Zero matters as evidence about the training path. It is not the polished public identity that readers usually mean when they say R1.

R1 added a multi-stage training process

DeepSeek-R1 added cold-start data and a multi-stage process involving reinforcement learning and supervised fine-tuning. The stated aim was to retain reasoning performance while improving readability and general behavior.

The full R1 model builds from the DeepSeek-V3 base architecture. V3 is a 671-billion-total-parameter mixture-of-experts model with 37 billion parameters activated per token, according to the V3 technical report.

This means full R1 should not be described as a 37-billion-parameter model. The 37 billion figure refers to parameters activated per token under the cited architecture.

flowchart TD
    A["DeepSeek-V3-Base"] --> B["R1-Zero research path"]
    A --> C["R1 multi-stage path"]
    C --> D["Full DeepSeek-R1"]
    D --> E["Distilled Qwen checkpoints"]
    D --> F["Distilled Llama checkpoints"]
    D --> G["Hosted service behavior"]
    E --> H["Possible third-party quantizations"]
    F --> H

Distilled models are separate checkpoints

The official R1 repository lists six dense distilled models at different sizes. Four use Qwen 2.5 bases. Two use Llama bases.

Distillation transfers behavior from a larger teacher into a smaller student during training. The smaller result is not the full mixture-of-experts model. It has different parameter count, memory, runtime, hardware, quality, failure behavior, and license chain.

A Qwen-based 32B distillation should be named as that artifact. Calling it simply "R1" makes a benchmark or deployment record impossible to reproduce.

Third-party conversions and quantizations create another identity. Lower-precision weights may reduce memory or improve speed while changing quality, numerical behavior, runtime compatibility, or security exposure.

The weights and the API are different products

Available weights let an organization select a checkpoint and runtime. That can improve visibility into the artifact and data path. It also creates responsibility for hardware, isolation, patching, scaling, security, monitoring, backups, incidents, and retirement.

DeepSeek's hosted service shifts some infrastructure work to the provider. It adds current model aliases, version drift, terms, data policy, rate limits, price, availability, retention, provider access, and service dependency.

The current API documentation controls present service claims. It should be checked immediately before implementation.

A prompt result from the website or API does not establish how a repository checkpoint behaves locally.

"Open source" needs an artifact

The R1 repository says its code and principal weights are available under the MIT License, with commercial use, modification, derivative works, and distillation allowed under the stated terms.

The same repository says the distilled checkpoints retain obligations from their Qwen or Llama base families.

That creates a license chain:

"Open weight" is often the more precise public term when discussing available model weights. A repository license does not automatically cover training data, hosted service, every dependency, or every use.

Author benchmarks belong to the release record

DeepSeek reports benchmark results for R1, R1-Zero, and distilled checkpoints in the repository and paper.

Those results are useful historical claims. They do not prove performance on another organization's tasks, language, data, prompt template, quantization, runtime, hardware, risk, latency, or cost.

Evaluation should preserve the exact artifact, source revision, date, license, quantization, tokenizer, runtime, hardware, prompt format, generation settings, tools, test set, scoring, repetitions, latency, tokens, cost, failures, and reviewers.

[[How to Evaluate an Open-Weight Reasoning Model]] provides the full evaluation frame.

Current use requires a current check

DeepSeek R1 is now a historical family with a maintained public repository. That does not mean every hosted alias or package remains unchanged.

Before use, verify the exact checkpoint or API identifier, lifecycle, revision, artifact hash, license chain, runtime support, context, tools, price, terms, data treatment, availability, safety behavior, and known limitations.

Then test the workload and deployment path you actually plan to operate.

This page was developed with AI assistance from the raw E055 transcript and linked primary sources, then structured for human model-identity, license, technical, currentness, security, and editorial review. It is not a license opinion, deployment approval, or model recommendation.