DeepSeek Open-Sources DSpark, Cutting AI Text Generation Time by Up to 85%

Per a new technical paper surfaced on Hacker News, Chinese AI lab DeepSeek has open-sourced DSpark — a suite of inference optimizations that accelerates language model text generation by 60 to 85 percent — and released the entire package for free.

This is not a model release. DSpark targets the inference layer, the machinery that converts a prompt into output tokens, and extracts dramatically more throughput from the same hardware. That distinction matters: inference costs are where hyperscalers and startups alike lose money at scale, and an 85% improvement in generation speed translates directly into lower cost per query. DeepSeek's own models have already matched OpenAI and Anthropic on standard benchmarks while running on considerably less compute; DSpark pushes that efficiency philosophy down into the infrastructure stack.

The move to open-source the optimization puts every major lab in an uncomfortable position. Proprietary inference stacks built over years of internal R&D are suddenly competing against a free alternative from a lab that has made a habit of giving away what others charge for. Startups that couldn't afford production-scale AI deployment are the immediate winners.

Whether DSpark's gains replicate across diverse production workloads is a question the community will answer fast — the code is live and anyone can run it now.

Read the full story at Hacker News