DeepSeek 把“印钞机”开源了：单卡每小时 38 美元，155 万亿 Token/ 天，推理框架正式放开
English headline: DeepSeek Open-Sources Its“Money Machine”: $38 Per GPU Hour, 155 Trillion Tokens a Day, and the Inference Stack Goes Public
10 月 20 日，DeepSeek 宣布把内部生产级推理框架 DeepInfer 全面开源。之前它已经陆续放出模型、TileLang 编译工具和推理算子库，这次直接把线上大规模推理服务的底座交了出来。
On October 20, DeepSeek announced that it is fully open-sourcing DeepInfer, its internal production-grade inference framework. It had previously released models, the TileLang compiler, and inference operator libraries; this time it is handing over the foundation that powers its large-scale online inference service.
这套框架已经支撑 DeepSeek V4.1 Flash 的高吞吐、低延迟线上服务。按官方披露，单一集群每天吞吐达到 155 万亿 Token，按谷价口径计算，单卡每小时收入约 38 美元。
The framework already powers DeepSeek V4.1 Flash with high throughput and low latency. According to official figures, a single cluster processes 155 trillion tokens per day, generating about 38 dollars per GPU hour under off-peak pricing.
DeepInfer 集成了高性能算子和深度优化的推理引擎，支持分布式推理、推测解码、SWA Bounded Replay、结构化输出等功能，并且同时兼容英伟达和华为昇腾平台。
DeepInfer bundles high-performance operators with a deeply optimized inference engine. It supports distributed inference, speculative decoding, SWA Bounded Replay, structured output, and runs on both NVIDIA and Huawei Ascend platforms.
同步开源的还有 deepseek-recipe，这是一套生产级 Token 编解码组件。它把官方 API 的请求处理、提示词渲染和输出格式转换逻辑完整放出来，让社区自建服务时能尽量对齐官方体验。
Also released is deepseek-recipe, a production-grade token encoding and decoding component. It exposes the official API’s request handling, prompt rendering, and output formatting logic, so community deployments can closely match the official service behavior.
DeepSeek 表示，模型决定能力边界，但推理框架决定服务成本。接下来，新模型的推理代码和性能优化会继续在 DeepInfer 里更新，目标是让大模型部署更透明、更便宜，也更容易被更多企业和开发者用起来。
DeepSeek says models define capability boundaries, but inference frameworks determine service cost. Going forward, inference code and performance improvements for new models will continue to land in DeepInfer, aiming to make large-model deployment more transparent, cheaper, and easier for enterprises and developers to adopt.
