DeepSeek Released R1, Its First Public Reasoning Model
By Jeremy ·
DeepSeek dropped R1 today alongside the full technical report and model weights, all under MIT license. Anyone can download, run, and build on it commercially without restriction.
What R1 Actually Is
R1 is a Mixture of Experts model with 671 billion total parameters, but only around 37 billion activate per token during inference. That architecture keeps compute costs manageable despite the overall parameter count.
The training approach is the more interesting part. DeepSeek trained an intermediate model called R1-Zero using pure reinforcement learning with no supervised fine-tuning at the start. R1-Zero developed chain-of-thought reasoning on its own through RL, without being shown examples of how to reason. The final R1 model adds a cold-start phase using a small set of curated long chain-of-thought examples before the RL training runs, which cleaned up some of the readability issues R1-Zero had.
Both models show their reasoning to users by default. You can follow the thinking process before the final answer, which is different from how most closed models present output.
DeepSeek also released distilled versions derived from the R1 reasoning patterns, ranging from 1.5 billion to 70 billion parameters. These run on consumer hardware and are already showing up in benchmarks ahead of models much larger than they are.
The Benchmark Numbers
On AIME 2024, R1 scored roughly 79.8% pass@1. On MATH-500, roughly 97.3% pass@1. On Codeforces, it landed in the 96th percentile range.
Those numbers are competitive with OpenAI's o1 on the same tasks, and in some cases ahead of it. DeepSeek's technical report covers the full benchmark breakdown if you want to read the methodology.
The comparison to o1 matters because o1 is what OpenAI positioned as its frontier reasoning model. DeepSeek published a model that sits in the same tier, open-sourced it, and put it out the same week.
Why the Cost Number Matters
DeepSeek reported training costs of approximately $5.9 million for R1. That figure covers the final training run, not all the compute spent on research and failed attempts along the way, so it's not a direct comparison to what OpenAI spent developing o1. The actual total investment is higher.
That said, the gap is still significant. US frontier labs have spent hundreds of millions of dollars on comparable model development. The training efficiency DeepSeek achieved using GRPO, a reinforcement learning algorithm that skips the need for a separate critic model, is part of why the cost is lower than expected.
Whether you think $5.9M is the full picture or not, the direction it points is clear: frontier reasoning capability is getting cheaper to replicate.
What Happened to US AI Stocks
Nvidia's stock dropped roughly 17% on the day R1 dropped. The reaction reflects a specific concern: if capable reasoning models can be trained cheaply and made freely available, the argument for spending billions on GPU clusters at scale gets harder to make.
The hardware scaling thesis has been the backbone of Nvidia's valuation run over the past two years. A model that performs at frontier levels for a fraction of the training cost is a direct challenge to that assumption, even if the picture is more complicated when you account for inference infrastructure and ongoing research compute.
The open weights mean any developer can run R1 locally, fine-tune it, or build products on top of it without paying for API access. That changes the competitive dynamics for closed model providers in a way that's difficult to walk back. The official response from US AI labs was to release more blog posts.