Deepseek R1: The Open Source Reasoning Model that turns the AI world upside down | 27.01.2025
In the international AI scene, there’s currently more excitement than ever before. While the U.S. announces multi-billion-dollar projects like “Project Stargate,” a new model called DeepSeek R1 is causing a stir in China. The unique aspect: it’s freely available, open source, and apparently powerful enough to go head-to-head with top language models from major companies like OpenAI—often at a fraction of the cost, or even for free. I like AI and have often written about Openai and other developments here in the blog. In this post, we’ll explore what DeepSeek R1 is, how it compares to other models, and the new possibilities it offers.
DeepSeek R1 is a reasoning model—a Large Language Model (LLM) specialized in chain-of-thought (CoT) reasoning. CoT models are designed for complex tasks such as math, advanced coding, or any type of problem requiring logical inference. What makes DeepSeek R1 especially interesting is that it’s completely open source. Developed by a Chinese company (or as a side project of a hedge fund), its aim is to create long-term social value rather than simply maximizing profit.
Performance Level
Initial tests and benchmarks show that R1 can match the performance of OpenAI’s O1 model—and even surpass it in some areas. AI experts like Dr. Jim Fan (Nvidia) and Emad Mustaq (Stability AI) praise the model, calling it a real milestone. Another remarkable point is the cost comparison: while using certain advanced reasoning features from OpenAI can be quite expensive, DeepSeek R1 is often either free or significantly cheaper.
A key distinguishing feature of DeepSeek R1 is its training method:
No Traditional Supervised Fine-Tuning Pipeline
Instead of using a dataset with step-by-step solutions (e.g., labeled by humans), DeepSeek R1 employs direct reinforcement learning.
Direct (or “Pure”) Reinforcement Learning
The model receives tasks without being shown how to solve them beforehand. Instead, R1 “tries out” different solutions, and the best ones receive rewards through a feedback system. This approach is somewhat akin to how humans learn by iterating on attempts until they get it right.
Thanks to this method, R1 has apparently developed its reasoning abilities in a way that, in many respects, mirrors human problem-solving. When you chat with R1, you can often follow its “chain of thought” in real time, providing unique insights into the AI’s internal thought process.
The quickest way to test R1 is via the official website deepseek.com. There, you’ll find a chatbot:
Activating the “DeepThink” Mode
This ensures that you’re using the R1 model instead of the default “DeepSeek V3,” making it better suited for complex questions and tasks.
Free & No Subscription
Registration is free, with no “Plus” plan. The thinking process is transparent, allowing you to see how the model arrives at its answers step by step.
For those who might be skeptical of using a Chinese website, or who simply want more independence, DeepSeek R1 can be run locally:
Smaller “Distilled” Models
For less powerful hardware, there are variants with 7, 32, or 70 billion parameters. These models are much more resource-friendly than the full version (at around 680+ billion parameters) but are also somewhat less powerful.
Software Tools
Tools like LM Studio, Ollama, or “Anything LLM” allow you to download and run the model on your own machine. The larger the model, the more powerful your hardware (RAM, GPU, CPU) needs to be.
Example: Counting Letters
A classic test for language models is counting specific letters in a word (e.g., “How many E’s are in Erdbeere?”). On high-end hardware running the full model, this task can be handled reliably. Smaller models may struggle or take longer to produce accurate results.
The release of DeepSeek R1 has caused an uproar, with some describing it as a “Sputnik moment.” Here we have a Chinese open-source model seemingly going toe-to-toe with expensive, closed-source American solutions. This development raises serious questions for commercial AI services and GPU manufacturers like Nvidia: if advanced AI models can run on relatively affordable hardware in the future, existing business models might be shaken.
OpenAI is working on its own reasoning models (like O1 or O1 mini) and offers additional features such as “Operator” (browser automation). However, the pressure from DeepSeek R1 is intense: R1 is not only open and transparent, but also less expensive—often even free. Sam Altman has responded on Twitter with somewhat cryptic messages, while investors appear uneasy.
One potential downside is censorship of certain topics. There are reports that R1 is heavily restricted—or doesn’t show its thought process at all—when prompted about sensitive issues (e.g., Taiwan or historical events like April 15, 1989). Since the model is open source, forks could emerge that remove these filters—a double-edged sword, as seen with so-called “liberated” versions of Meta’s Lama.
DeepSeek R1 shows that modern AI need not be the exclusive domain of large corporations. The era of requiring thousands of GPUs to run a robust model seems to be fading.
Opportunities for Developers and Businesses
With its MIT license and open interfaces, startups and established companies can integrate R1 into their products without hefty fees for access or API calls.
AI Hype and Next Steps
Some claim that AI is reaching its peak, whereas others see DeepSeek R1 as just the start of a new era. The field is evolving rapidly, and what is celebrated as a breakthrough today could be outdone by newer models tomorrow.
DeepSeek R1 delivers on what many have been hoping for: a free, high-performance reasoning model capable of tackling complex tasks—like advanced math, coding, and analytical problems—on par with, or even surpassing, some closed-source offerings. Its open-source nature and the possibility of local hosting not only increase transparency, but also give developers and companies more freedom.
Whether R1 signals a new era of freely available AI powerhouses or stands as a standout success in a fast-moving industry remains to be seen. One thing is certain: with this model, China has handed the global AI community a remarkable gift—one that could reshape the balance of power in the AI world. Anyone who wants to stay relevant in AI should keep a close eye on DeepSeek R1 and give it a try, as it’s rarely been this easy and cost-effective to run a capable reasoning model on your own hardware.