DeepSeek has introduced two new language models — V4-Pro and V4-Flash. Both are already available for free in the DeepSeek chatbot, and their open weights have been published on the Hugging Face platform. According to the company’s own tests, the flagship model outperforms most open competitors and approaches the level of ChatGPT and Gemini.
DeepSeek has released new models and they’re top-tier
What DeepSeek V4-Pro and V4-Flash Can Do
V4-Pro is DeepSeek’s flagship model. Its total size is 1.6 trillion parameters, but only 49 billion are activated per request. This architecture (called Mixture of Experts) allows the model to be powerful without spending all resources on every response.
V4-Flash is a lighter version with 248 billion parameters, of which 13 billion are used per generation. According to DeepSeek, it delivers results on par with the flagship model for simple tasks, but responds faster and is cheaper to operate.
Both models feature a context window of 1 million tokens — this is the volume of text the model can consider when generating a response. For comparison, an average-length book fits into roughly 100,000–150,000 tokens. Both also support a “reasoning” mode, where the model breaks down a task step by step, and a standard mode for quick answers.
How DeepSeek V4-Pro Differs from ChatGPT and Gemini
DeepSeek provides results from its own benchmarks, and it’s worth noting that these are the company’s own data, not independent tests. Nevertheless, the numbers are noteworthy.
According to DeepSeek, V4-Pro surpasses all existing open models in math, programming, and natural science tests. In terms of “general world knowledge,” it falls short only of Google’s Gemini 3.1 Pro. V4-Pro promises to be even stronger than previous versions.
DeepSeek is shown in blue on the chart. Keep in mind that all the other models are paid, while DeepSeek is free
The company also claims V4-Pro leads in benchmarks for so-called agentic programming — this is when AI doesn’t just answer a question but carries out a chain of actions: writes code, runs it, checks the result, and fixes errors. The model supports integration with Claude Code, OpenClaw, and OpenCode tools, which developers use for writing programs with AI assistance.
How to Enable DeepSeek V4 in the Chatbot for Free
The main advantage of DeepSeek for regular users is accessibility. The DeepSeek chatbot works for free, and both new models are already available in its “Expert” mode. For comparison, access to ChatGPT’s most powerful models requires a Plus or Pro subscription.
If you use AI for everyday tasks — writing text, understanding complex topics, getting coding advice, or analyzing documents — DeepSeek V4 can cover most of these needs without a subscription. The one-million-token context window allows you to load an entire long document into the model and ask questions about it.
That said, it’s worth noting that DeepSeek is a Chinese company, and its servers are located in China. For some users, this may be an important factor. No special steps are needed to enable the new models in DeepSeek. Here’s how it works:
Simply go to the website, select the mode, and start using it
- Open the DeepSeek chatbot at chat.deepseek.com or in the mobile app
- Switch to “Expert” mode for the Pro model and “Fast” mode for Flash
Preview versions of both models are available right now.
API access is promised by the company on April 24, 2026 — this is important for developers who want to integrate the models into their products. API pricing has not been announced yet. The model weights have also been published on Hugging Face — this means that technically advanced users and companies can run the models on their own servers.
Which DeepSeek Model to Choose
V4-Flash is a sensible choice for most tasks: quick answers to questions, help with texts, simple code, and information summarization. It responds faster and, according to DeepSeek, doesn’t fall behind the flagship model in simple scenarios.
V4-Pro is worth choosing for complex tasks: deep analysis, mathematics, programming, working with long documents, and chains of reasoning. If you need AI as a working tool for development or research, the flagship model will deliver noticeably better results.
A free alternative to ChatGPT for those who don’t need image generation
For Apple device owners, this means DeepSeek remains one of the few fully free alternatives to ChatGPT that works in the Safari browser and through the app. Given that Apple Intelligence is still limited in capabilities and unavailable in some regions, DeepSeek fills the need for a powerful AI assistant without subscriptions or regional restrictions.
If you’re paying for ChatGPT Plus and using it for simple tasks, now is a good time to try DeepSeek V4 and decide whether you still need that subscription.
