Via cnet.com
The Chinese AI lab's latest model brings massive scale and agentic capabilities to its web platform, mobile app, and API under an MIT license
DeepSeek has officially launched its V4 Pro model, dropping one of the largest open-weight AI systems ever released into general availability. The model packs 1.6 trillion total parameters and deploys 49 billion of them per token, a scale that puts it in direct competition with the proprietary heavyweights from OpenAI, Google, and Anthropic.
The difference: anyone can use it. V4 Pro ships under an MIT license, which is about as permissive as software licensing gets.
What’s under the hood #
DeepSeek-V4-Pro uses a Mixture-of-Experts (MoE) architecture. Think of it like a company with 1.6 trillion employees where only 49 billion show up to work on any given task. The rest stay dormant until their specific expertise is needed.
The model supports a 1 million-token context window.
DeepSeek has integrated V4 Pro across its full product surface: web platform, mobile app, and API. The company has also added Responses API support and Codex integration, both of which are geared toward agentic capabilities. Architectural improvements include hybrid attention mechanisms, which help the model handle both short, focused queries and sprawling long-context tasks.
From preview to production #
V4 Pro isn’t entirely new. DeepSeek first released a preview version on April 24, 2026, giving developers an early look at the model’s capabilities before committing to a full production rollout.
Alongside the preview, DeepSeek also released a lighter variant called V4 Flash, which carries 284 billion total parameters with 13 billion active parameters per token.
The general availability update for V4 Pro emphasizes improvements specifically tuned for production environments, making the model viable for always-on, high-stakes deployment.
The open-weight gambit #
The MIT license is the quiet headline here. While companies like OpenAI and Anthropic keep their model weights locked behind API paywalls, DeepSeek is making its most powerful model freely available for modification, fine-tuning, and commercial deployment.
For developers and startups, open weights mean the ability to customize the model for specific domains without paying per-token fees to a cloud provider. A healthcare company can fine-tune V4 Pro on medical literature. A fintech firm can specialize it for regulatory analysis. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our