DeepSeek released 3 new open-source V4 models

DeepSeek released preview versions of its open-source V4-Pro and V4-Flash models.

Both are mixture-of-experts models with a default 1M-token context window and new attention/compression techniques.

The API supports reasoning and non-reasoning modes, including a reasoning_effort setting for different use cases.

DeepSeek says the models post benchmark results close to leading proprietary systems, including in long-context and coding tasks.

The launch raises the bar for open-weight AI and adds pressure on closed-source frontier labs.