In a landmark moment for the AI community, the release of DeepSeek R1 sent waves through the tech world, not for disrupting markets, but for boldly redefining the potential of open-source AI. This seemingly modest reasoning model didn’t just push boundaries, it reimagined them.
DeepSeek R1 is a pivotal development that challenges long-standing assumptions about the exclusivity of advanced AI. By delivering sophisticated reasoning capabilities at a fraction of the traditional cost, it dismantles the notion that powerful AI must remain confined behind proprietary walls. Instead, it paves the way for a future where cutting-edge intelligence is accessible, affordable and inclusive, transforming how developers, researchers and communities interact with AI.
Understanding AI: Non-reasoning vs. reasoning models
Before diving into DeepSeek's innovations, it's worth understanding a crucial distinction in today's AI landscape.
Most large language models (LLMs) we interact with daily, including earlier versions of ChatGPT and similar tools, are primarily "non-reasoning" models. They're extraordinarily good at pattern recognition and language prediction but cannot methodically work through complex problems step by step.
Reasoning models, by contrast, can break down complicated tasks into logical sequences—much closer to how humans’ approach problem-solving. This capability has been the crown jewel of proprietary systems like OpenAI's o1 model, which required massive computational resources and investment to develop.
These advanced reasoning capabilities would remain the exclusive domain of deep-pocketed tech giants for the foreseeable future but DeepSeek R1 shattered that assumption overnight.
Building with constraints
DeepSeek's journey is particularly fascinating because it began with significant constraints. While US tech companies had access to Nvidia's most powerful H100 GPUs, DeepSeek had to make do with the H800—a "nerfed" version delivering only about 70% of the H100's performance for LLMs due to export restrictions.
This limitation might have spelled doom for less innovative teams. For DeepSeek, it became the catalyst for reimagining how AI models could be built more efficiently.
The technical breakthroughs behind DeepSeek R1
The ‘mixture-of-experts’ advantage
Rather than running every input through the entire neural network (as traditional "dense" models do), DeepSeek adopted a Mixture-of-Experts (MoE) architecture. Think of it as assembling a specialized team for each task rather than consulting the entire company.
When processing language, the model dynamically routes inputs to only the most relevant "expert" sub-networks. By activating just a fraction of the model's parameters for any given task, MoE dramatically reduces computational demands while maintaining or improving performance.
Smart precision choices
DeepSeek further optimized resource usage through intelligent precision selection. For most operations, the model uses FP8 (8-bit floating point) calculations, essentially working with less precise numbers where they don't impact quality. This nearly doubles computational speed while cutting memory usage.
The model switches to higher-precision formats for the most sensitive calculations where precision matters. This balanced approach ensures accuracy isn't sacrificed for efficiency.
The Group Relative Policy Optimization (GRPO) innovation
DeepSeek's most groundbreaking contribution came with its novel approach to reinforcement learning called Group Relative Policy Optimization (GRPO).
Traditional reinforcement learning evaluates each potential response in isolation. Instead, GRPO looks at groups of possible reactions, establishing a baseline and comparing individual answers against it. This approach proved particularly effective for enhancing reasoning capabilities while minimizing training costs. Here’s how it works in simple terms:
- Group evaluation: When the model receives a prompt, it generates several possible responses. Instead of judging each answer independently, GRPO looks at all the responses as a group.
- Baseline and rewards: It calculates an average reward (a baseline) from these responses. Each response is then compared to this baseline. The rewards are based on two key factors:
- Accuracy: Whether the final answer is correct according to predefined rules.
- Format: Whether the response follows a specific structure, like using special tags (e.g., and ) to show the reasoning process clearly.
- Policy update: The model uses the differences between each response’s reward and the group baseline to update its strategy, or policy, for generating future outputs. This means that over time, the model learns to produce correct and well-structured answers.
This approach was first applied to the DeepSeek-V3 base, creating DeepSeek-R1-Zero. Initially, the team recognized that while DeepSeek-R1-Zero had shown promising reasoning capabilities, its outputs suffered from poor readability and language mixing. To address these shortcomings, they began by conducting a supervised learning phase using the DeepSeek-V3-Base model. This phase involved fine-tuning a carefully curated training dataset comprised of high-quality, human-friendly Chain-of-Thought (CoT) examples.
Once the model was primed with this enhanced readability, it was introduced to the Group Relative Policy Optimization (GRPO) technique. This reinforcement learning phase was pivotal in further refining the model’s reasoning abilities. Here, the team added a language consistency reward. This new reward component penalized outputs that mixed languages, ensuring that the CoT remained consistent with the target language.
In summary, the training process was cleverly split into two phases:
- First, a supervised learning phase using high-quality, structured examples to establish clear reasoning patterns and readability
- Then, GRPO-based reinforcement learning with an added language consistency rule
This DeepSeek R1 model—released under an MIT license—proved that top-tier AI reasoning could be achieved at a fraction of the cost through more innovative architecture and training methods. This shattered the narrative that only the most prominent players could compete in cutting-edge AI, exposing the vulnerability of the economic moats protecting massive AI investments.
Economics of AI innovation
We're entering a new phase of AI development where clever engineering and algorithm design might matter more than raw computing power and capital. This does not mean computing is irrelevant, but DeepSeek's innovations demonstrate that the relationship between resources invested and capabilities achieved isn't linear.
This creates a more complex landscape for investors to navigate. The questions shift from "Who has the most resources?" to "Who is using those resources most intelligently?" Companies that had positioned themselves as AI leaders based primarily on their ability to outspend competitors suddenly look vulnerable.
Barriers to breakthroughs
What we are witnessing in the current AI landscape is not merely a technical evolution, but a fundamental reimagining of the economics of AI. This disruption is paving the way for broader access to advanced AI capabilities for both consumers and businesses. As barriers to entry continue to fall, competition intensifies, driving prices downward and accelerating the democratization of technology.
The key challenge lies in discerning which companies are best positioned to succeed in this new paradigm—and which may struggle to justify their AI investments in a world where the principle of “more is better” no longer guarantees an edge.
