DeepSeek R1 Explained

What it is

DeepSeek R1 is an open-weights reasoning large language model that works through problems step by step, showing a long chain of thought before answering. It specializes in mathematics, competitive programming, science, and complex logic. DeepSeek-AI, a Hangzhou-based company spun out of the quantitative hedge fund High-Flyer and founded by Liang Wenfeng, released it on January 20, 2025 alongside its technical report.

The release was a genuine shock to the industry. Here was a model matching OpenAI's o1 on reasoning benchmarks, with MIT-licensed weights anyone could download, at API prices far below the competition. Nvidia lost roughly $600 billion in market value on January 27, 2025, a 17 percent single-day drop, as investors repriced assumptions about how much compute frontier AI requires. The DeepSeek app briefly topped the US App Store's free charts. US officials began raising privacy and national security concerns within days.

How it works

R1 is a 671-billion-parameter mixture-of-experts model with 37 billion active parameters per token, built on the DeepSeek-V3 base. Only a fraction of the parameters engage for any given token, which is how it stays efficient despite its size.

Its training broke with convention. Instead of imitating human-written reasoning traces, the team used large-scale reinforcement learning with a method called Group Relative Policy Optimization, which rewards verifiable correct answers without needing a separate value model. An early variant, R1-Zero, used pure reinforcement learning with no fine-tuning at all and still climbed from 15.6 to 71.0 percent on the AIME 2024 math benchmark, with reasoning behaviors like self-reflection and backtracking emerging on their own.

The final R1 used a four-stage pipeline. Cold-start fine-tuning gave it readable output, a reasoning-oriented reinforcement learning stage built the capability, rejection-sampled fine-tuning on 800,000 curated examples polished it, and a final alignment stage cleaned up what remained. That pipeline fixed R1-Zero's poor readability and its tendency to mix languages mid-answer. Users see the reasoning in exposed think tags, which made the model's process as interesting as its answers.

Six smaller distilled models, from 1.5 to 70 billion parameters, were released alongside it, with the 32B and 70B versions claimed to perform on par with o1-mini. Distillation let the reasoning ability run on far smaller hardware.

Key details

The context window is 128,000 tokens, and the modality is text only. On paper-reported benchmarks against OpenAI's o1, R1 scored 79.8 percent on AIME 2024 against o1's 79.2, 97.3 on MATH-500, reached the 96.3rd percentile on Codeforces, and resolved 49.2 percent of SWE-bench Verified tasks. The weights, code, and report are all MIT licensed and widely mirrored on Hugging Face and third-party inference providers.

Launch-era API pricing was $0.55 per million input tokens and $2.19 per million output, compared with $15 and $60 for OpenAI's o1. That thirty-fold gap was the number that traveled around the industry. Rates change over time, so verify current pricing before relying on these figures.

Limitations

The documented censorship is the serious caveat. Testing by PromptFoo across 1,360 sensitive prompts found an 85 percent refusal rate on China-sensitive topics; the hosted app refuses subjects like Tiananmen Square outright. An arXiv study found censorship patterns persisting even in locally run weights, with a second model layer deleting or parroting official narratives on sensitive subjects. This is not a quirk of the hosted app; it appears baked into the model itself.

There were also early quality issues. R1-Zero mixed languages and produced messy output, problems the later pipeline stages addressed. And the market-shock narrative deserves a footnote: efficient training does not eliminate the demand for compute, it redirects it, a dynamic economists have understood since Jevons described it for coal in the 1860s. R1 is a remarkable engineering achievement with a political filter baked in, and anyone deploying it should understand both halves of that sentence.

The aftershock

The market reaction to R1 was immediate and went beyond the usual tech-stock volatility. The January 27 selloff erased hundreds of billions in value from chipmakers on a single thesis: if frontier reasoning could be trained this efficiently, the demand projections for AI compute might be overstated. Whether that thesis was right is still debated. Cheaper capability tends to expand total usage rather than shrink it, but the scare itself showed how much of the industry's valuation rested on the assumption that scale was the only moat.

The competitive response was fast. Reasoning models went from a niche to the default: every major lab shipped models with explicit chain-of-thought within months, and "reasoning" became a standard model category rather than a research novelty. The distilled R1 variants, small enough to run on modest hardware, spread reasoning capability far beyond the data center. API pricing across the industry came under pressure as DeepSeek's rates reset expectations for what inference should cost.

For the open-weight community, R1 was the second watershed after Llama, and arguably the bigger one. Llama had shown that open models could be good; R1 showed they could be frontier. The censorship baked into the weights tempered the celebration and sparked serious work on understanding how political filtering survives fine-tuning and distillation. That research continues, and it matters for anyone building on open models from any source: weights carry the values of their trainers, and those values are not always visible until someone goes looking.