DeepSeek V4 Preview: 1.6T parameters, 1M context and MIT weights in two sizes
In one sentence DeepSeek releases V4 Preview with open weights under the MIT license: V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active). 1M-token context; with hybrid CSA+HCA attention, at 1M tokens V4-Pro uses 10% of V3.2's KV cache.
On April 24, 2026, fifteen months after DeepSeek R1, DeepSeek publishes V4 Preview: downloadable weights under the MIT license, two sizes.
- DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion active per token (it's a Mixture-of-Experts).
- DeepSeek-V4-Flash: 284 billion total, 13 billion active. For users who want speed and economy.
Both handle 1 million tokens of context, now the default across all official DeepSeek services. The technical novelty: hybrid attention (Compressed Sparse Attention + Heavily Compressed Attention) that, at 1M tokens, cuts the KV cache to 10% of DeepSeek-V3.2's. Translation: much less GPU memory for the same input length.
The models are on Hugging Face and in the DeepSeek API, which is compatible with both the OpenAI and the Anthropic formats.
Companies
DeepSeek
Tools
DeepSeek V4 Preview, DeepSeek-V4-Pro, DeepSeek-V4-Flash
Tags
Sources