DeepSeek Launches V4 with 1.6 Trillion Parameters and 1 Million Token Context

By Hector Herrera | April 24, 2026 | Science

DeepSeek released preview versions of its V4 flagship model today, delivering the most capable open-source AI system yet released by any lab, East or West. The release comes exactly one year after DeepSeek's V3 upended Silicon Valley's assumptions about AI development, and arrives on the same day the Trump administration announced a crackdown on Chinese firms it says are exploiting U.S.-developed AI models.

What DeepSeek Released

CNBC reports that DeepSeek shipped two V4 variants simultaneously:

• V4-Pro: 1.6 trillion total parameters, with 49 billion active per token. Uses a Mixture-of-Experts (MoE) architecture — a design that activates only a portion of the model's total parameters for any given input, making massive models computationally practical to run.
• V4-Flash: 284 billion total parameters, 13 billion active. A lighter variant built for lower-latency applications and cheaper deployment.

Both models ship with a 1 million token context window — enough to process roughly 750,000 words in a single prompt. For comparison, most commercial models cap at 128,000 to 200,000 tokens.

The models also introduce a Hybrid Attention Architecture, a new design DeepSeek says dramatically improves long-context memory — the model's ability to recall and reason over information from early in a very long document or conversation.

Both variants are available immediately via the DeepSeek API. Both are also released as open-source, meaning the weights are publicly downloadable for local deployment.

Why the Context Window Is the Key Number

A 1 million token context window changes what's practically possible with AI. It means feeding an entire legal case file, a full year of a company's internal emails, or a complete codebase into a single prompt. Current commercial models can approximate that with chunking workarounds — but native support at this scale is a different capability.

The MoE architecture explains how V4-Pro achieves 1.6 trillion parameters without requiring the GPU cluster to process all 1.6 trillion simultaneously. At inference time, only 49 billion parameters activate per token — a design that makes the model expensive to train but relatively efficient to run at scale.

One Year Later

DeepSeek's V3 release in late 2025 was the inflection point. It matched top U.S. models on standard benchmarks at a fraction of the reported training cost. That single release triggered a brief Nvidia stock selloff and forced a public reckoning in Silicon Valley about whether American labs still held a meaningful capability advantage.

V4 arrives with no apologies. 1.6 trillion parameters makes V4-Pro one of the largest Mixture-of-Experts models ever released publicly — open-weight or otherwise.

The open-source aspect is not incidental. Unlike GPT-4o or Claude, which are accessible only through controlled APIs, V4's open weights mean any developer, company, or government can download and run the model independently, modify it, and build on it — with no ongoing relationship with DeepSeek required.

What to Watch

Independent benchmark results on Chatbot Arena (LMSYS), MMLU, and coding evaluations will arrive within days. Those scores determine whether V4-Pro competes with or surpasses closed frontier models from OpenAI, Anthropic, and Google at their current capability levels.

The policy response will follow close behind. The Trump administration's announcement today — targeting Chinese companies exploiting U.S. AI models — signals that government attention to Chinese AI development is intensifying. A credible 1.6 trillion parameter open-source model from a Chinese lab will accelerate that debate.

Source: CNBC