DeepSeek open-sources DeepInfer, the engine behind its API

On October 20 DeepSeek released DeepInfer under Apache 2.0, the inference framework powering its public API. It supports only DeepSeek-V4.1-Flash (763B parameters, 1M-token context) and requires Nvidia SM100 GPUs (B200/GB200) or Huawei Ascend 950 NPUs.

• DeepInfer reached about 22,300 GitHub stars by October 25
• The smallest efficient cluster is 128 GPUs: 64 for prefill and 64 for decode
• Renting that cluster on CoreWeave costs $880,000 per month
• Self-hosting beats the API only above 76% sustained GPU utilization