DeepSeek r1: The free AI model that challenges market leaders

DeepSeek R1, a new open source LLM developed in China, is attracting unprecedented interest. This revolutionary model not only technically rivals giants such as GPT-4 or Claude 3.5, it offers a bold vision of AI: transparent, freely accessible and resource-efficient.

Listen to the AI podcast :

So what are the technical features of DeepSeek ? What challenges does it give to tech giants and the chip industry ? And above all, how could it redefine the future of AI ?

Nvidia stock drop and DeepSeek r1 impact

On January 27, 2025, Nvidia took a spectacular tumble on the stock market, with the stock plunging 13% at the opening and reaching a total drop of over $400 billion.

At the day’s close, the stock was down 16.86%, the biggest drop since 2020, and Nvidia lost its place as the world’s largest market capitalization.

DeepSeek R1, causing a shockwave

Comparable performances

DeepSeek R1 offers comparable performance to the world’s best chatbots, including those developed by giants like OpenAI.

Lower Cost

DeepSeek R1 significantly reduces the cost of entry into the field of artificial intelligence compared to the billions invested by Nvidia and competing companies.

Resource efficiency

DeepSeek R1 runs on a more limited number of advanced Nvidia chips than its competitors, demonstrating resource efficiency in the development of a high-performance model.

Open Source

Available as open source, DeepSeek R1 accelerates adoption and development, allowing any company or developer to benefit from its performance without additional investment.

Popularity

The DeepSeek app has become the most downloaded free iPhone app in the U.S., showing its growing popularity among users.

Implications for the AI Industry

The emergence of DeepSeek R1 calls into question several aspects of the artificial intelligence industry:

Massive investments in question

Colossal investments such as those by Microsoft ($80 billion) and Meta ($60-65 billion) in AI infrastructure may seem pointless if a high-performance model can be developed with fewer resources.

American technological advantage challenged

American technological supremacy is challenged by a Chinese start-up’s ability to develop such a successful model, underscoring the importance of talent and global prospects in AI.

Technical capabilities of DeepSeek r1

A state-of-the-art architecture

DeepSeek R1 is based on an advanced architecture derived from its predecessor, DeepSeek V3, a model composed of 671 billion parameters.

What sets DeepSeek apart is its innovative approach: a mixture of experts system.

In contrast to conventional architectures activating all parameters simultaneously, DeepSeek only activates 37 billion parameters per token, optimizing efficiency while minimizing resource consumption.

DeepSeek R1 goes one step further with:

• Unsupervised reinforcement learning: The model evaluates itself by comparing its answers to reference solutions.
• “Chain of Thought” inference (CoT): He “thinks out loud”, breaking down complex problems into clear steps while self-correcting in real time.
  
  This improves the reliability of generated answers, particularly in demanding fields such as programming or mathematics.

Outstanding efficiency

DeepSeek R1 impresses not only with its technical capabilities, but also with its optimization of resources:

• Training performed with 2.78 million GPU hours H800 (designed to be exportable to China), far less powerful than the A100 or H100 used by OpenAI.
• Comparison : GPT-4 would have required around 60 million GPU hours, i.e. 95 % faster training for DeepSeek V3.

These results demonstrate that training high-performance models is not necessarily synonymous with massive raw power, but can be achieved through optimized resource management.

High-level performance

Despite modest hardware, DeepSeek R1 rivals or surpasses GPT-4 and Claude 3.5 in critical areas such as :

• Mathematics: Complex resolutions achieved with precision.
• Programming: Exceptional performance on GitHub Copilot.
• Language tests: A fluent, relevant ability to understand and respond in multiple languages.

Accessibility: AI for everyone

Contrary to other closed models, DeepSeek is open source. This means that its source code and weights are publicly accessible.

Furthermore, distilled versions, optimized to run on consumer hardware, offer local integration possibilities, making it a truly democratic technology.

At Anthem Creation we’ve added Deepseek’s latest model to LM Studio, and the results are convincing… If you’re looking for a alternative to ChatGPT that’s free, install an AI locally. We recommend using LM Studio or Jan.ai

Why DeepSeek redefines the rules

DeepSeek is distinguished by major strengths that place it as a redoubtable challenger in the field of AI.

Cost reduction

DeepSeek’s advances enable decreasing by dozens the costs involved in training and deploying AI models:

• Ideal for startups and researchers with limited budgets.
• Greater accessibility for previously excluded sectors (education, SMEs).

Democratizing AI

The open source aspect and the ability to run DeepSeek on common hardware favor large community adoption:

• Developers worldwide can adapt the template to their needs.
• Enhanced collaboration, conducive to innovation.

Transparency and control

In contrast to proprietary models, DeepSeek offers better transparency. Researchers can analyze how it works in detail, which helps reduce bias and promote responsible use.

Adaptability and distilled models

Distilled versions allow the template to be run on standard PCs or even mobile devices. This adaptability makes it an essential tool for a variety of uses : embedded AI, home automation systems, etc.

To better understand model distillation, read our article: GPT model distillation via API: OpenAI’s new game-changing breakthrough

The impact of DeepSeek on the GPU chip industry

DeepSeek could transform the current ecosystem, particularly in the semiconductor field. Here are the main implications:

1. Reduced demand for high-end GPUs : Models like DeepSeek show that modest GPUs can suffice, which could threaten Nvidia’s business model.
2. Transition to specialized chips : An opportunity for the development of hardware designed for distilled AI.
3. Lower data center costs : Thanks to drive energy efficiency, companies could reduce their operational expenses.

DeepSeek: A paradigm shift

DeepSeek embodies a new era for AI : one of accessibility, efficiency and collaborative innovation.

By democratizing advanced tools and shaking up established business models, it paves the way for a future where AI serves everyone, not just tech giants.

FAQ : Everything you need to know about DeepSeek

1. Can DeepSeek run offline ?
Yes, thanks to the distilled templates, DeepSeek can be run offline via tools like LM Studio.

2. What’s the difference between DeepSeek R1 and DeepSeek V3 ?
DeepSeek R1 introduces unsupervised reinforcement learning and Chain of Thought inference, two major innovations.

3. What hardware is required to run DeepSeek ?
Mid-range GPUs (e.g. RTX 3060) are sufficient for distilled models.

4. Is DeepSeek really comparable to GPT-4 ?
For some benchmarks, yes. However, GPT-4 remains slightly superior in specific complex tasks.

5. How to access DeepSeek ?
Via their website, mobile app, or locally with LM Studio.

6. What is a distilled version ?
An optimized, slimmed-down version of an AI model to reduce computing power requirements.

7. Why is DeepSeek open source ?
To foster collaboration and democratize access to AI.

8. Does DeepSeek consume a lot of energy ?
No, it’s designed to be extremely energy and resource efficient.

9. Can DeepSeek be adapted to specific needs ?
Absolutely, it can be customized thanks to its open source approach.

10. What are the challenges of DeepSeek ?
Risks include limitations in certain complex tasks and geopolitical challenges related to its provenance.