DeepSeek-R1

tl;dr: open reasoning model family from DeepSeek, MIT-licensed at the weights level, with the full 671B MoE plus six distilled sizes. Default ollama run deepseek-r1 lands on the 0528-Qwen3-8B distill — fast, decent at math, fits on a laptop.

What it is

deepseek-r1 is the umbrella tag Ollama ships for the DeepSeek-R1 release. The base weights are an MoE reasoning model trained with RL, but the family also includes distillates fine-tuned from Qwen and Llama checkpoints. All sizes expose reasoning traces (<think>...</think> blocks) and tool use.

The page on ollama.com lists tools and thinking as capabilities, 35 tags total, 92.6M cumulative pulls — second in the Ollama library behind Llama 3.1.

Pull command

From the ollama.com detail page:

ollama run deepseek-r1          # defaults to deepseek-r1:8b (Qwen3-8B distill)
ollama run deepseek-r1:1.5b
ollama run deepseek-r1:7b
ollama run deepseek-r1:14b
ollama run deepseek-r1:32b
ollama run deepseek-r1:70b
ollama run deepseek-r1:671b     # full MoE

Pull deepseek-r1 once and Ollama caches the 8B distill. To upgrade from a previous install: ollama pull deepseek-r1.

Sizes and hardware

Context is 128K for every size except the 671B, which is 160K.

License

From the ollama.com page:

• Base DeepSeek-R1 weights: MIT License. Commercial use, modifications, derivative works (including distillation for training other LLMs) all allowed.
• Qwen distillates (1.5B / 7B / 14B / 32B): derived from Qwen-2.5, originally Apache 2.0.
• Llama 8B distill: derived from Llama 3.1-8B-Base, inherits the Llama 3.1 community license.
• Llama 70B distill: derived from Llama 3.3-70B-Instruct, inherits the Llama 3.3 community license.

The full license texts are not linked from the Ollama page. If you’re shipping commercial output from the Llama distillates, check the upstream community license separately — Meta’s commercial restrictions still apply there.

What the README says

The ollama.com README describes the 0528 update as a “minor version upgrade” that “significantly improved reasoning and inference capabilities,” with the 8B distill and the 671B MoE both bumped. DeepSeek claims performance approaching O3 and Gemini 2.5 Pro on math, programming, and general logic benchmarks.

The page also lays out the distillation story: DeepSeek generated reasoning traces from the full R1 model, then fine-tuned smaller dense checkpoints on those traces. Result: the 7B and 8B distillates outperform what RL-from-scratch gets you at the same size.

Gotchas

• The default deepseek-r1 is the 8B distill, not the headline 671B. If you want the MoE, you have to specify deepseek-r1:671b — and you need real hardware.
• Reasoning output is verbose. The <think> blocks eat tokens and slow throughput. Budget for them.
• Distillates inherit upstream licenses. The 70B is a Llama 3.3 distill, so Meta’s commercial restrictions apply on top of the base DeepSeek MIT terms.
• 671B is 404GB on disk. Plan storage before pulling.
• Reasoning quality varies more by size than you’d expect — the 1.5B is fine for code reading but loses coherence on multi-step math.

Related: Llama 3.1 for the upstream weights behind the 8B and 70B distillates, Ollama for the runtime.

Untested here — facts from the ollama.com /library/deepseek-r1 page fetched 2026-09-10.