DeepSeek / DeepSeek-V

DeepSeek V4-Pro Weights

Open-weight foundation model packaging in final preparations following initial API serving evaluation benchmarks.

Expected✓ Corroborated

Source checked:

Launch window

2026 · Q4

No verified launch time. We do not infer a countdown from an estimated window.

Open-Weight Foundation Milestone

DeepSeek V4-Pro Weights represents the anticipated open-source release of DeepSeek’s high-efficiency Mixture-of-Experts architecture for global researchers and commercial developers.

Architectural Enhancements

Featuring refined Multi-head Latent Attention (MLA) and enhanced sparse routing, V4-Pro offers near-frontier reasoning performance with drastically reduced memory footprints during inference.

Availability Horizon

Expected in the fourth quarter of 2026 (2026-Q4) following initial cloud evaluation phases, verified via authenticated GitHub repository preparations.

Latest documented updates

  1. Open weight packaging in preparation for wider developer community release.

    Read source

Release timeline

  1. Open-weight milestone anticipated following initial API serving benchmark evaluations.

    Read source

Sources & verification

These sources support the historical milestone. Verification of an announcement does not imply current product availability.

  1. DeepSeek open-source repositories and model cards Official source

Related releases

Delayed

DeepSeek

DeepSeek-R2

Second-generation pure reinforcement learning reasoning model postponed for specialized domestic hardware adaptation.

Date to be announced✓ Sourced
Confirmed

Alibaba Cloud

Qwen 4

Multi-trillion parameter foundation model confirmed by Alibaba Cloud at Apsara Conference for 2027 launch.

2027✓ Sourced