DeepSeek V4: Chinese AI model with 1.6 trillion parameters defies American giants

One year after shaking the market with a competitive and open-source model, DeepSeek features the V4 Preview, its newest artificial intelligence system. There are two versions — the standard and the Pro, which boasts 1.6 trillion parameters (49 billion active per inference), becoming the largest open weight model ever released. The company claims that V4 competes directly with closed systems such as OpenAI GPT-4, Google Gemini and Anthropic Claude.

How does DeepSeek V4 with 1.6 trillion parameters compare to market leaders?

According to DeepSeek's statement, V4 achieves comparable or superior performance in mathematical reasoning, coding, and language comprehension benchmarks. Although the complete results of independent benchmarks have not yet been released, the previous one already puts the model at the same level as American leaders. Meanwhile, the AI arms race accelerates with new OpenAI, DeepSeek and Anthropic models, and DeepSeek bets on opening as a strategic differential.

Why is DeepSeek V4's open weight model a differential?

Unlike virtually all direct competitors, who keep their models closed and charge for API, DeepSeek provides V4 weights for free. This allows developers, startups and researchers to download, modify and run the model on their own infrastructure, without relying on expensive services. The approach can accelerate innovation in AI, especially in countries with restrictions on access to American platforms. This movement reinforces the trend pointed out in articles such as the hallucinating rhythm of AI releases generates FOMO in users, says product head of Claude Code.

What to expect from the full launch of DeepSeek V4?

The preview already impresses, but the final version of DeepSeek V4 promises more optimizations and the publication of a detailed paper with official benchmarks. The company did not disclose an exact date, but the final model should further expand the advantage of being the largest open model available. Experts see the movement as an attempt to democratize access to state-of-the-art AI, while Google and OpenAI intensify their own releases, such as the $750 million invested by Google Cloud in agency AI (as reported in $750 million: Google Cloud Heavy Bet on AI Genetics with partners.

The original source, The Verge, highlights that DeepSeek has again managed to surprise the industry, proving that open models can rival the best proprietary systems.

Common Questions about DeepSeek V4

Is DeepSeek V4 free?

Yes, the preview of the model is available for free download and testing, but the company has not yet announced prices for commercial use in production.

How many parameters does the model have?

The V4 Pro has 1.6 trillion total parameters, with 49 billion active per inference. The standard version had no detailed specifications, but is smaller.

Can the model be used commercially?

Yes, because it is open weights, developers can integrate it into products as long as they respect the license terms still to be disclosed by DeepSeek.