Overview

DeepSeek released a preview of the V4 series in April 2026 with two mixture-of-experts models. A hybrid of Compressed Sparse Attention and Heavily Compressed Attention makes long contexts cheaper: at 1M tokens, DeepSeek reports that V4-Pro needs 27% of the per-token compute and 10% of the KV cache of DeepSeek-V3.2. Both models were pre-trained on more than 32T tokens and use the Muon optimizer.

Highlights

  • V4-Pro: 1.6T total, 49B active parameters
  • V4-Flash: 284B total, 13B active parameters
  • One-million-token context window
  • Manifold-Constrained Hyper-Connections (mHC) for stable training
  • Base and instruct weights on Hugging Face under MIT

Typical uses

Long-document analysis, large codebases, agents and reasoning, mostly through the API or large self-hosted deployments.

Things to consider

  • Released as a preview; DeepSeek later issued updated checkpoints such as V4-Flash-0731.
  • Even V4-Flash needs server-class hardware.

Specifications are taken from the model card on Hugging Face and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.