DeepSeek Quietly Launched V4 Pro. Currently Only Available via API

The new model appears without a PR campaign

DeepSeek has released the V4 Pro 0813 model without putting together a traditional landing page or official announcement. Developers only noticed the update when the model appeared on the OpenRouter aggregation platform, through which it is currently available for API calls. The open weights for the 1.7 trillion parameter, 893 GB model are already available for download on Hugging Face.

Three layers of reasoning produce different outputs

An interesting detail of the deployment on OpenRouter is the availability of three different levels of internal logical reasoning: low, medium, and high. Initial independent tests demonstrate that the model produces visually and stylistically distinct results depending on the selected reasoning level, a characteristic not typically this pronounced in competing platforms. For example, when generating an SVG pelican, details such as the drawing style or the color of a bicycle vary.

A shift in distribution models for Chinese labs

The missing PR campaign illustrates a contrasting approach to releases. While Western labs build massive announcements around every minor version update, Chinese players increasingly opt for silent drops through community channels and aggregators. Benchmarks for the new model were initially circulated in a closed WeChat group before users copy-pasted them to Reddit, where moderators deleted them for lacking context, before they permanently landed as ASCII tables on Hacker News.

The community discovers the model without marketing

API availability on OpenRouter is crucial for initial testing. However, the real test for the V4 Pro model was the release of its open weights, which the community had been waiting for. The publication of the weights demonstrates that Chinese labs rely more on developers discovering and adopting a high-quality model independently of any absent PR communication.

Lilith's verdict

The era of every minor model update requiring a grandiose blog post is over. WeChat leaks and OpenRouter availability are more than enough to reach the exact crowd that will actually stress-test it.