❯ DeepSeek open-sources DeepInfer, the core engine behind large-scale V4.1-Flash inference

Production experience reaches developersDeepSeek open-sourced DeepInfer on October 20 to serve trained models efficiently. The release supports DeepSeek V4.1-Flash on NVIDIA and Huawei Ascend platforms under Apache 2.0. Core features include distributed inference, speculative decoding, asynchronous scheduling and structured outputs.

Less repeated computation and idle hardwareIts deployment documentation describes separately scheduling input processing and output generation, while retaining reusable context calculations. DeepSeek reports that a production cluster serves about 155 trillion tokens daily, generating approximately $38 per GPU-hour at off-peak prices.

The release has defined boundariesCache management and load-balancing code tied closely to internal infrastructure are not yet fully included. Developers must still supply components for a complete online service. The accompanying deepseek-recipe, which converts API requests, encodes tokens and parses outputs, uses the MIT license.

Revenue is an operational metricRevenue per GPU-hour describes the business served by a unit of compute; it does not establish corporate gross margins. Training, networking, storage, staff and service redundancy are not fully captured in that figure. Teams hosting models can instead reuse the published methods and code, then calculate economics for their own traffic, latency targets and hardware.

▮ SIGNALOpen inference engines let third parties reuse system-level optimization, making deployment economics increasingly sensitive to caching, scheduling and real request patterns.