DeepInfer

DeepInfer is software for running DeepSeek models to produce outputs; this open-source release provides offline inference with core features for serving at scale. ^[1] It supports DeepSeek-V4.1-Flash and includes features such as optimized kernels (specialized computation routines), parallelism (splitting work across devices), and structured outputs (constraining generation to a regex, JSON schema, or grammar). ^[2a][2b][2c] A supported use case is offline inference with the default DeepSeek-V4.1-Flash checkpoint, as shown by the offline inference example. ^[3]