DeepSeek just made a bold statement in the artificial intelligence industry with the launch of its V4 model, a giant of approximately 1 trillion parameters (with Mixture of Experts and 37 billion active per token) that runs on Huawei Ascend 950PR chips instead of the ubiquitous Nvidia GPUs. Published under an Apache 2.0 license, this model represents a milestone in the democratization of AI and has profound implications for developers, companies, and governments in Latin America. In this article we analyze its architecture, benchmarks, the geopolitical context surrounding it, and what it means for those of us who work with AI in the region.

The architecture behind DeepSeek V4

DeepSeek V4 uses a Mixture of Experts (MoE) architecture that reaches approximately 1 trillion total parameters, but only activates 37 billion for each token processed. This technique makes it possible to achieve performance comparable to much larger dense models, while keeping inference computational costs significantly lower.

Key specifications:

• Total parameters: ~1 trillion.

• Active parameters per token: 37B, which keeps latency low.

• Context window: 1 million tokens, allowing extensive documents to be processed without fragmentation.

• Modalities: Text, image, and code. Natively multimodal.

• License: Apache 2.0, the most permissive open source license.

• Training hardware: Huawei Ascend 950PR chips.

The Huawei factor: training without Nvidia chips

Perhaps the most disruptive aspect of DeepSeek V4 is not its size but the hardware it was trained on. While practically all frontier models in the world are trained on Nvidia GPUs (A100, H100, H200, B200), DeepSeek managed to train V4 using Huawei Ascend 950PR chips, designed and manufactured in China.

According to Reuters, the semiconductor export restrictions imposed by the United States on China have driven the accelerated development of domestic AI chips. DeepSeek V4 is the most compelling proof that this strategy is yielding results. The Ascend 950PR, although it doesn't reach the specifications of Nvidia's H100 in individual synthetic benchmarks, proved to be sufficient to train a frontier-class model when deployed in massive clusters optimized with proprietary software.

This has enormous implications for the global industry: if top-tier models can be trained without Nvidia chips, the de facto monopoly that Nvidia maintains over AI training starts to crack. For LATAM, where accessing the latest-generation Nvidia GPU clusters is extremely costly, hardware alternatives could reduce the barriers to entry in the medium term.

Benchmarks: frontier-level performance

DeepSeek V4 demonstrates competitive performance with the best models on the market across a variety of standardized benchmarks:

While benchmarks don't tell the whole story (real performance depends on the specific use case), DeepSeek V4's numbers firmly place it in the frontier category, competing on equal terms with models from OpenAI, Anthropic, and Google.

The 1-million-token context window

With 1 million tokens of context, DeepSeek V4 can process extremely long documents without needing chunking or RAG techniques. This opens up concrete possibilities:

• Legal analysis: You can load complete contracts, regulations, and relevant case law in a single query.

• Code auditing: An entire code repository can be analyzed at once, identifying bugs, vulnerabilities, and refactoring opportunities.

• Academic research: Multiple papers can be compared and synthesized in a single interaction.

• Data processing: Complete datasets can be analyzed without needing pre-processing scripts.

Apache 2.0: real open source

DeepSeek chose to publish V4 under the Apache 2.0 license, the most permissive open source license for commercial use. This means that any company, developer, or researcher can:

• Download and use the model at no cost.

• Modify it and fine-tune it for specific use cases.

• Integrate it into commercial products without restrictions.

• Redistribute it (maintaining the attribution).

This contrasts with models like GPT-4 (proprietary, only accessible via API), Claude (proprietary), and Llama (Meta's custom license with restrictions). DeepSeek's decision to use Apache 2.0 is a strategic bet on massive adoption: by eliminating legal and economic barriers, the model can become the foundation of an enormous ecosystem of applications and fine-tunes.

Multimodal capabilities

DeepSeek V4 is natively multimodal, which means it can process and generate content in multiple formats:

• Text: Comprehension and generation of text in multiple languages, including Spanish.

• Images: Image analysis, description, OCR, and answering questions about visual content.

• Code: Generation, analysis, debugging, and refactoring of code in more than 50 programming languages.

Implications for Latin America

The launch of DeepSeek V4 has several concrete implications for the region:

Reduction of AI costs

Being open source and optimized for alternative hardware, DeepSeek V4 can be run on more accessible infrastructure. Companies in Argentina, Mexico, Colombia, and Brazil that today pay high costs for proprietary model APIs could migrate to their own DeepSeek V4 instance, reducing costs by up to 80% in some use cases.

Data sovereignty

Running a model locally means the data never leaves your infrastructure. For regulated industries (banking, health, government), this is a key enabler. A hospital in Buenos Aires could use DeepSeek V4 to analyze medical records without sending sensitive data to servers in another country.

Geopolitical diversification

According to MIT Technology Review, the concentration of frontier AI in a handful of US companies is a systemic risk. The emergence of viable alternatives like DeepSeek V4 diversifies the ecosystem and reduces dependence on single providers, something especially relevant for LATAM governments that are defining their national AI strategies.

Opportunities for developers

With a frontier model accessible under Apache 2.0, developers in LATAM can build globally competitive AI products without needing the multimillion-dollar budgets that training your own model requires. Fine-tuning DeepSeek V4 for a specific use case (legal, medical, educational) is a concrete business opportunity.

To understand how DeepSeek V4 positions itself against other frontier models like GPT-5, we recommend reading our article about the launch of GPT-5 and its comparison with Google's Gemini.

How to start using DeepSeek V4

If you want to try DeepSeek V4, you have several options:

• Official API: DeepSeek offers an API with competitive prices. You can register at platform.deepseek.com and start making calls in minutes.

• Local download: The model weights are available on Hugging Face. You need significant hardware to run the full model (multiple GPUs with at least 128 GB of total VRAM), but quantized versions can run on more modest configurations.

• Cloud providers: Platforms like Together AI, Fireworks, and Groq offer access to DeepSeek V4 with competitive per-token prices.

• Ollama: For smaller quantized versions, Ollama allows you to run the model locally with a single command.

Recommended Coderhouse course

If you want to understand in depth how these models work and apply them in real projects, these Coderhouse training programs are ideal:

• AI Fundamentals Course: Understand the fundamental concepts behind models like DeepSeek V4: neural networks, transformers, fine-tuning, prompt engineering, and more. Ideal if you want a solid foundation for working with AI.

• AI Automation Career: Learn to integrate AI models like DeepSeek into real automation flows. From connecting with APIs to deploying complete solutions in production.

Frequently asked questions

Is DeepSeek V4 really better than GPT-4o?

It depends on the benchmark and the use case. In standardized benchmarks like MMLU-Pro and HumanEval, DeepSeek V4 shows competitive and in some cases superior results. However, real performance varies according to the specific task, the language, and the prompt. What's most significant is that an open source model reaches this level of performance, because it democratizes access to top-tier AI.

Can I run DeepSeek V4 on my computer?

The full model requires specialized hardware (multiple high-end GPUs). However, quantized versions (Q4, Q8) can run on more modest configurations. A Q4 quantized version of the base model can run on a GPU with 24 GB of VRAM, although with a reduction in performance. For professional use, the most practical approach is to access it via API or use a cloud provider.

What does it mean that it runs on Huawei chips instead of Nvidia?

It means China demonstrated the ability to train frontier AI models without depending on Nvidia's technology, which dominates more than 90% of the GPU market for AI. For the global industry, this introduces competition in AI hardware and could reduce costs in the long term. For LATAM, it opens the possibility of accessing alternative and potentially more economical AI hardware.

Is it safe to use a Chinese AI model?

Being open source under Apache 2.0, the model's code and weights are completely auditable. Anyone can inspect exactly what the model does. If you run it on your own infrastructure, the data doesn't leave your server. The security risks are the same as with any AI model (hallucinations, biases, malicious use) and are not inherent to the model's geographic origin.

Does Apache 2.0 mean I can use it for anything?

Apache 2.0 is one of the most permissive open source licenses. You can use DeepSeek V4 for personal, commercial, academic, governmental, and any other purpose. The only obligations are to maintain the copyright notice and the license in distributions. You don't have to open source your own code if you use DeepSeek V4 as a component.