DeepSeek open-sources DeepInfer for large-scale model inference
The framework runs V4.1 Flash on NVIDIA and Huawei hardware. Its offline inference core is public, while parts of DeepSeek’s production serving system remain outside the release.
THE DEEPSEEK DESK
DeepSeek.
Model releases, research and developer updates from DeepSeek.
13 STORIESDATES REFER TO ORIGINAL ANNOUNCEMENTS
DeepSeek opens its search engine to developers through new APIs
Search and Fetch give developers another route to bring web information into their AI applications.
DeepSeek updates V4.1 Flash for coding and agent work
V4.1-Flash-1017 reaches the app, website and API, with DeepSeek reporting shorter task runtimes and stronger results.
DeepSeek introduces V4.1-Flash with native visual understanding
The September release changes the architecture and API model name, while a later notice preserves access to V4-Pro.
DeepSeek opens an experimental vision model for agents
V4-Flash-Vision-Exp brought mixed text and image requests to the API in August; the model has since been superseded.
DeepSeek moves V4-Pro into general availability
The August release added reasoning controls and native Responses support, alongside a new peak and off-peak pricing structure.
DeepSeek’s July Flash update targeted coding agents
A new post-training run reached the V4-Flash API without changing the model’s size or architecture.
DeepSeek’s V4 preview made million-token context standard
The April release introduced Pro and Flash variants, open weights and updated API model names.
DualPath research tackles the storage cost of agent memory
The DeepSeek-listed systems paper adds a second route for loading cached context during multi-turn inference.
DeepSeek-OCR 2 explores a different reading order for images
The January research replaces a rigid visual scan with token ordering guided by image meaning.
Engram research gives language models a separate lookup memory
A January paper investigates how static memory and neural computation can share a model’s capacity.
DeepSeek’s mHC paper revisits how model layers connect
The year-end research proposes constraints intended to keep expanded residual connections stable at scale.
DeepSeek V3.2 brought reasoning into tool use
The December release paired a general model with a separate high-compute research variant.
↗
No DeepSeek stories in this edition yet.
This desk will feature source-linked coverage of DeepSeek.
Browse all stories