DeepSeek V4 Expected in April — 1 Trillion Parameters, Native Multimodal

DeepSeek V4 is expected to launch in April 2026, alongside Tencent's new Hunyuan model, according to Chinese tech outlet Whale Lab reported by Dataconomy. The model has been anticipated since mid-February, with multiple projected release windows passing without a public launch.

What We Know About V4

DeepSeek V4 is a ~1 trillion parameter Mixture-of-Experts model with only ~37B active parameters per token — meaning it can match the performance of much larger models while running on significantly less compute. The architecture includes a 1M-token context window powered by Engram conditional memory, a technology published on January 13 that enables efficient retrieval from extremely long contexts.

The model targets native multimodal generation: text, image, and video from a single architecture.

The Benchmark Claims

Leaked benchmarks claim 90% HumanEval and 80%+ SWE-bench Verified — which would match Claude Opus 4.6. These numbers are unverified and should be treated with appropriate skepticism until independent testing confirms them.

The Geopolitical Context

DeepSeek V4 is being optimized for domestic Chinese AI chips through partnerships with Huawei and Cambricon. This directly responds to US export controls on advanced semiconductors and aligns with China's push for AI hardware independence.

Meanwhile, OpenAI, Anthropic, and Google are cooperating to prevent model distillation — the technique DeepSeek previously used to train competitive models from Western frontier model outputs.

Our Take

If DeepSeek V4 delivers anywhere near its leaked benchmarks at the efficiency its architecture suggests, it will be the most cost-effective frontier model available. The 37B active parameters make it dramatically cheaper to run than Western alternatives. But "leaked benchmarks" from an unreleased model deserve exactly as much credibility as that phrase implies. Wait for the release.