DeepSeek Releases V4 Preview: Largest Open-Weights Model at 1.6T Parameters Under MIT License

DeepSeek released two preview models from their V4 series: DeepSeek-V4-Pro (1.6T total / 49B active parameters, Mixture-of-Experts) and DeepSeek-V4-Flash (284B total / 13B active). Both support 1M token context windows and are released under the MIT license, making them available for unrestricted commercial use and self-hosting. V4-Pro is now the largest open-weights model available, surpassing Kimi K2.6 (1.1T) and the prior DeepSeek V3.2 (685B).

The combination of frontier-class scale, open weights, and MIT licensing continues to challenge closed-model economics. The V4-Flash variant (13B active parameters) is particularly interesting for inference-cost-sensitive deployments. Security-relevant implication: unrestricted access to very large open-weights models increases capability availability for both defensive tooling and adversarial AI-assisted attack research.