DeepSeek has launched preview versions of its flagship V4 model today, introducing a 1.6 trillion parameter architecture with a one-million-token context window and open-source weights, one year after its V3 release rattled global markets and forced a rethink of AI development economics.
The timing is deliberate and sends a clear message to the global tech industry. Exactly one year after DeepSeek-V3 triggered a $600 billion single-day wipeout in Nvidia's market capitalization by proving that frontier AI performance could be achieved at a fraction of Western development costs, the Hangzhou lab led by CEO Liang Wenfeng is back with a bold new offering. The architecture is substantially more ambitious, pushing the boundaries of what open-source artificial intelligence can actually do. The release comes in two distinct variants designed for entirely different scales of operation. First, there is DeepSeek-V4-Pro, carrying 1.6 trillion total parameters with 49 billion activated per token to handle complex, heavy-duty reasoning tasks. This mixture of experts approach ensures that the model only uses the computing power necessary for the specific job at hand, keeping operational costs surprisingly low. Second, the company is deploying DeepSeek-V4-Flash, a leaner 284 billion parameter model with 13 billion activated per token, built specifically for rapid-fire deployments where speed and efficiency are the top priorities. Both models support one million token context windows as standard. This means developers can feed entire codebases, extensive financial documents, or whole novel series into the model in a single prompt without losing context or suffering from the degraded recall that often plagues smaller context models. Furthermore, both versions are open-sourced immediately, giving global developers commercial and research access from day one without waiting for restricted API tiers.
The benchmark claims are striking and deserve close attention from anyone tracking the AI industry. According to DeepSeek's official release notes, V4-Pro achieves state-of-the-art performance among open-source models in agentic coding, mathematical reasoning, and world knowledge. It trails only Google's Gemini 3.1 Pro in the knowledge category while rivaling the world's top closed-source models across math and STEM benchmarks. This directly challenges the assumption that proprietary systems will always maintain a comfortable lead over their open-source competitors. The open-source community now has a tool that can stand toe-to-toe with heavily funded proprietary models from major tech giants. In practical terms, this kind of mathematical and coding capability makes the model an incredibly powerful tool for enterprise automation, complex software engineering, and advanced scientific research. It essentially gives smaller startups and independent developers access to cognitive tools that were previously locked behind expensive corporate paywalls. A three-person startup can now run complex agentic workflows and perform deep data analysis at the same level as a massive enterprise. As TechCrunch confirmed this morning, the company says both preview versions are already being adopted by major open-source repositories and enterprise partners eager to integrate the new capabilities. This rapid adoption signals a growing trust in DeepSeek's ability to consistently deliver reliable, high-performing infrastructure that pushes the entire industry forward.
Also read: AI will almost certainly create the world's first trillionaire and the race is already underway • Intel's agentic AI bet sends SanDisk surging as the CPU quietly reclaims its throne • China's crackdown on US investment in AI firms signals a new front in the tech cold war
