DeepSeek has open-sourced DeepInfer, the internal inference framework that powers its V4.1-Flash model — the company’s fastest and cheapest AI model. Released under the Apache License 2.0, DeepInfer is the serving software that lets a trained model generate answers quickly at scale, and it’s the same engine DeepSeek runs in production.
The framework builds on vLLM, the open-source serving engine widely used across the industry, but adds DeepSeek’s own optimized compute kernels, distributed processing across multiple chips, and faster text-generation techniques. Those are the tricks behind the famously low prices of DeepSeek’s hosted API — and now anyone running the models on their own hardware can match that performance instead of being limited to the company’s servers.
DeepSeek is also releasing deepseek-recipe, a companion tool that replicates the token processing of the official DeepSeek API so self-hosted deployments behave identically to the real thing. Perhaps more significantly, the company says code for future models will ship as open source as well, extending the commitment beyond this single release.
The move piles pricing pressure on Western providers like ChatGPT‘s OpenAI and Claude‘s Anthropic, whose customers now have a clearer path to self-hosting a competitive model stack at a fraction of the cost. For developers comparing open options, DeepSeek and Gemini remain the key names to watch in the cheap-model race.
Source: The Decoder
