DeepSeek-V4-Pro

Open Weights

DeepSeek

DeepSeek's frontier open-weights MoE model (April 2026): 1.6T total / 49B active parameters with a 1M-token context window and a hybrid attention mechanism (Compressed Sparse Attention + Heavily Compressed Attention) for efficient long context. MIT-licensed and competitive with top closed models on math, STEM, and coding.

Strengths

Rivals top closed models on Math/STEM/Coding while fully open-weight
1M-token context with highly efficient hybrid attention (CSA + HCA)
Permissive MIT license - free for commercial self-hosting
Far lower long-context inference cost than DeepSeek-V3.2

Caveats

1.6T total parameters require a large multi-GPU deployment to self-host
Preview release - behavior may change before stable

Vendor: DeepSeek
Context Window: 1,000,000 tokens
License: MIT
Release Date: 2026-04-24
Modalities: text
image

Capabilities

Vision

Audio

Video

Tool Use

Resources

API Documentation Hugging Face

Pricing

Open weights under the MIT License - free to self-host; managed API/pricing via DeepSeek and other providers.