DeepSeek R1 Guide: Reasoning, Uses and Local Access

DeepSeek R1 is an open-weight reasoning model family released in 2025. The official repository includes the full 671B R1 and R1-Zero models plus distilled checkpoints from 1.5B to 70B. R1 remains relevant for research and local use, but it is not the current first-party API model family in July 2026.

R1 in one minute

• R1 was trained from a DeepSeek V3 base for reasoning.
• R1-Zero explored reinforcement learning without an initial supervised fine-tuning stage.
• Distilled models transfer R1-generated reasoning data into smaller Qwen and Llama bases.
• The repository lists 128K context for full R1 and R1-Zero.
• Local practicality depends on checkpoint size, quantisation and hardware.

R1, R1-Zero and R1-0528

R1-Zero is primarily a research result showing reasoning behaviours emerging from reinforcement learning. Its outputs can be less readable or consistent for general users. R1 adds training stages intended to improve usefulness and presentation.

R1-0528 is a later update announced in May 2025. Runtime catalogues may use the bare deepseek-r1 tag for a newer 8B distilled checkpoint, while :671b refers to the full model. Always inspect the exact runtime manifest instead of assuming a tag means the original release.

Distilled model family

The official R1 repository lists Qwen-based 1.5B, 7B, 14B and 32B variants and Llama-based 8B and 70B variants. Later packaging includes an updated Qwen3 8B distill. Distillation means the smaller model was fine-tuned on reasoning data generated by R1; it is not the same architecture or capability as the full 671B model.

Smaller models are easier to run but can lose reliability, depth and instruction following. Evaluate the exact checkpoint on your tasks. A result about full R1 cannot be assigned to an 8B quantised distill.

Useful R1 workloads

R1 is designed for tasks that benefit from multi-step reasoning: mathematics, algorithm design, code debugging, constraint problems and structured analysis. It can also help critique a proposed solution by checking assumptions and testing edge cases.

Reasoning models still make errors. Request an auditable answer with equations, code, test cases or cited evidence. Do not equate a long reasoning trace with correctness. For factual research, current sources matter more than the appearance of careful thought.

Hosted and API status

At launch, R1 was accessible through a deepseek-reasoner API alias and a DeepThink chat mode. DeepSeek later changed the underlying alias through V3.x releases and scheduled both legacy aliases for retirement after 2026-07-24.

The current API lists V4 Flash and V4 Pro with thinking controls. New applications should follow the current API guide, not copy R1-era request examples. If you need the R1 checkpoint specifically, use a verified local or third-party deployment.

Run R1 locally with Ollama

Ollama’s official library lists:

ollama run deepseek-r1:1.5b
ollama run deepseek-r1:7b
ollama run deepseek-r1:8b
ollama run deepseek-r1:14b
ollama run deepseek-r1:32b
ollama run deepseek-r1:70b
ollama run deepseek-r1:671b

The default tag can change. Run ollama show deepseek-r1 after pulling and record the digest. The complete Ollama guide covers installation and local API security.

Hardware expectations

Model size labels are parameter counts, not direct RAM requirements. Quantisation lowers memory use at a possible quality cost. Context length and concurrent users add memory pressure. CPU-only inference may work for smaller models but can be slow; GPU support varies by platform and runtime.

Avoid universal claims such as “runs on any laptop.” Download size alone can exceed available storage, and full R1 is a specialist deployment. Start with a small checkpoint and a short context, then measure speed and quality on the target machine.

Licensing

The R1 repository states that its released weights use the MIT licence and permit commercial use, modification and distillation. It also notes that distilled Qwen and Llama models derive from base models with their own licences.

Read all applicable terms for the exact checkpoint. A third-party quantisation should identify its source and licence. Model permission does not remove obligations related to training data, output use, privacy or local law.

Prompting R1

Give a clear problem, input constraints and required final format. Ask for the final answer plus checkable evidence, not an unbounded essay. For code, require tests; for mathematics, require substitution or a separate verification; for decisions, require alternatives and failure conditions.

Limit output and stop repetitive reasoning. If a smaller distill loops, restate the target format and break the problem into stages. The prompt guide provides original patterns.

Limitations

R1 can hallucinate facts, mishandle ambiguous requirements, produce insecure code and overuse verbose reasoning. Distilled variants can behave differently from the full model. Local releases lack hosted search unless the surrounding application supplies it.

Benchmarks in the model card are developer-reported under specified settings. They are useful evidence about the release, not a universal ranking or a prediction of your production task.

Conclusion

DeepSeek R1 is best treated as a model family with one large reasoning release, a research variant and several distinct distills. Choose by exact checkpoint, licence, hardware and task—not by the R1 name alone. Use V4 IDs for current first-party API work.