DeepSeek V4 Preview: 1.6T parameters, 1M context, open weight in two sizes

In one sentence DeepSeek releases V4 Preview as open source: V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active). Native 1M-token context, hybrid CSA+HCA attention cutting KV cache by 90%.

Verified Official source

ShareLinkedIn X

Exactly one year after the DeepSeek R1 "Sputnik moment" (January 2025), which surprised the market with frontier performance at 1/20th the cost of Western competitors, DeepSeek returns on April 24, 2026 with V4 Preview. Open source, downloadable weights, two sizes.

DeepSeek-V4-Pro: 1.6 trillion total parameters (49 billion active per token, it's a Mixture-of-Experts). The largest open-weight model published so far.
DeepSeek-V4-Flash: 284 billion total, 13 billion active. For users who want speed + economy.

Both support a native 1-million-token context. The real technical novelty: they use a hybrid attention architecture (CSA + HCA, internal DeepSeek terms) that cuts KV cache by 90% vs V3.2. Translation: they run on much less GPU memory for the same input length.

For researchers, enterprise IT self-hosting, countries that don't want to depend on US models: V4 is an open frontier-grade option. Available on Hugging Face, DeepSeek API (compatible with both OpenAI and Anthropic API formats).