DeepSeek V4 Ships a 1.6 Trillion Parameter Open Model

DeepSeek shipped two new open-weights models on April 24, both released under the permissive MIT licence. The headline configuration is V4-Pro, a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active per token, making it the largest open-weights language model anyone has published. The lighter V4-Flash carries 284 billion total parameters with 13 billion active. Both ship a one-million-token context window, with output capped at 384,000 tokens.

The model weights are live on Hugging Face, free for download and commercial use. That detail matters. A 1.6T parameter model under MIT licence is, by some measures, the most capable openly licensed model anywhere.

The numbers behind the numbers

What makes V4 interesting is not the parameter count, which is a slightly old way to talk about model capability, but the efficiency claims. Winbuzzer reports that V4 incorporates a hybrid attention mechanism combining Compressed Sparse Attention and Heavily Compressed Attention. In the one-million-token context setting, V4-Pro requires only 27 percent of the inference compute and 10 percent of the KV cache that the previous V3.2 generation needed.

That matters because long-context inference is what kills serving costs. Most labs have hit a wall where doubling the context window roughly doubles the price per query. If DeepSeek's compression numbers replicate in the wild, the company has built a model that runs whole codebases or long discovery documents through itself without the typical cost penalty.

The wider picture

This is the second open-weights frontier model to land in the past month. Moonshot's Kimi K2.6 arrived on April 20, also at trillion-scale, also focused on long-horizon agentic coding. Together with Alibaba's Qwen 3.5 family and Z.ai's GLM-5.1, the picture is now fairly clear: the open-source frontier is increasingly being defined by Chinese labs.

The cost gap is the part that should worry incumbents. Independent comparisons put V4-Pro inference at roughly one-twentieth the cost of running comparable Western frontier models for similar output quality on coding benchmarks. That is the kind of gap that changes procurement conversations, especially for cost-conscious enterprise buyers who do not need the absolute frontier.