The code is available under the Apache 2.0 license and currently supports one model, DeepSeek-V4.1-Flash, loading the updated 1017 version by default.
According to the report, DeepInfer supports Nvidia SM100 Blackwell GPUs and Huawei Ascend 950 NPUs, but not Nvidia Hopper or older architectures. Its reproducibility features work only on Nvidia hardware. DeepSeek describes an efficient minimum deployment as 64 GPUs for prefill plus 64 GPUs for decoding.
The framework uses DSpark speculative decoding, where a draft model proposes five tokens and a confidence head selects candidates for verification. DeepSeek estimates this raises throughput by 4% to 7% at production load. It also says a separate technique can nearly halve total prefill computation. Published benchmarks were run on Nvidia hardware and are DeepSeek's own figures.
The release does not include DeepSeek's key-value cache management or load-balancing code, which the company says is coupled to internal infrastructure. It does not establish independent performance results, broad multi-model support, or a small-scale deployment path.