4. Pretraining – Teaching the Model to Predict Language

At the heart of every LLM is a Transformer architecture, a deep learning framework that has revolutionized natural language processing (NLP). Instead of reading words sequentially like older models, Transformers use self-attention mechanisms, allowing the model to weigh the importance of different words in a sentence regardless of their position.

Example of Self-Attention:

In the sentence “The bank approved the loan”, the word “bank” refers to a financial institution.

In “The boat reached the riverbank”, “bank” has a completely different meaning.

Transformers analyse context, enabling them to determine the correct interpretation.

During pretraining, the model learns to predict missing words based on its context. Given the phrase:

“The cat sat on the [MASK].”

The model might predict “mat” based on prior knowledge of similar phrases in its training data.

This phase requires massive computational power—typically thousands of GPUs or TPUs—making it one of the most expensive steps in LLM development.

5. Fine-Tuning – Aligning the Model with Human Intentions

Once pretrained, the model understands language, but it still lacks task-specific abilities or human alignment. This is where fine-tuning comes in.

One of the most effective fine-tuning methods is Reinforcement Learning from Human Feedback (RLHF):

• Human annotators rank multiple model-generated responses.
• A reward model is trained to score responses based on quality and accuracy.
• The main LLM is further optimized using reinforcement learning techniques like Proximal Policy Optimization (PPO) to improve future outputs.

This process helps models become more helpful, honest, and safe, but it still requires significant human oversight and labelled data, making it labour-intensive.

6. Optimization – Reducing Model Size and Speeding Up Performance

Pretrained and fine-tuned LLMs are often too large and inefficient for real-world applications. Optimization techniques help reduce computational costs while maintaining performance:

Pruning – Removes unnecessary neurons or parameters

Quantization – Converts high-precision floating-point calculations to lower-bit values, such as 8-bit integers, for faster inference

Distillation – Trains a smaller “student” model to replicate the behaviour of a larger “teacher” model

These methods allow companies to deploy models more efficiently, reducing server costs and improving response times.

7. Deployment – Making LLMs Accessible to Users

Once optimized, LLMs are deployed through:

Cloud-based APIs (e.g., OpenAI, Anthropic, Google Gemini)

On-premises enterprise solutions (for companies needing data security)

Edge computing (running models on consumer devices to reduce cloud dependency)

At this stage, the model becomes accessible to developers, businesses, and end-users, powering everything from chatbots to coding assistants and enterprise automation tools.

What DeepSeek Did Differently? 

DeepSeek didn’t just refine this pipeline—it reimagined it. Here’s how:

1. Pure Reinforcement Learning Approach

Unlike traditional LLMs that heavily depend on supervised learning and human-labelled data, DeepSeek’s R1 model uses a reinforcement learning-first approach.

This means the model teaches itself, optimizing outputs through trial and error rather than relying heavily on human feedback. In other words, the model chooses to “reward” or “penalise” itself based on the outputs it generates during the RL process.

2. Mixture of Experts (MoE) Architecture

DeepSeek utilizes a Mixture of Experts (MoE) framework, which consists of 671 billion parameters but activates only 37 billion at a time, reducing computation costs while maintaining state-of-the-art performance and the R1 model can specialize in different tasks dynamically rather than treating all inputs equally.

The MoE framework is like having a team of specialised geniuses, each tackling different parts of a problem. Instead of one massive model doing all the work, MoE smartly deploys the best experts for each task, slashing computational costs and supercharging efficiency. During training, these experts are exposed to diverse datasets, making them versatile problem-solvers. A gating mechanism decides which experts to activate, and when multiple experts are activated, their outputs are seamlessly combined, boosting the model’s overall performance.

3. Cost-Effective Training with Minimal Compute

But what truly sets DeepSeek apart from giants like Google and Mistral (who already use MoE architecture in some of their models) is their ingenious quantization technique. By using 8-bit numbers for most model weights and switching to 32-bit only for complex calculations, DeepSeek cracked the code for training on less powerful NVIDIA hardware.

Consequently, DeepSeek signalled a shift in AI economics, proving that cutting-edge LLMs can be developed without requiring astronomical compute budgets. The future of LLMs isn’t about building bigger models—it’s about making them smarter and more efficient.

4. Open-Source Accessibility

Unlike OpenAI and Anthropic, DeepSeek released its model under an MIT open-source license, allowing anyone to use and modify it, accelerating AI research and development and opening the door to smaller companies and researchers to build on cutting-edge AI without needing massive funding.

Why DeepSeek Caused a Disruption 

As mentioned earlier, the disruption caused by the R1 model wiped almost $1 trillion off the stock market in a single day. This seismic reaction wasn’t just about DeepSeek itself—it was about what it represented. The AI industry has long operated under the assumption that state-of-the-art models require billion-dollar budgets, massive datasets, and the backing of Silicon Valley’s most powerful firms. DeepSeek shattered that illusion. It proved that a cutting-edge LLM could be developed with fewer resources, in the face of economic sanctions, and with a fraction of the compute traditionally thought necessary.