After DeepSeek, StartLux strikes again: open-sources the first decision model inference engine, inference up to 5.3x faster

What happened: StartLux open-sourced Prism, described as the first inference engine built specifically for decision models, two days after DeepSeek open-sourced its DeepInfer framework. Key numbers: Prism took 19 ms versus 53.1 ms for the prior setup on a single short question (2.8x faster), and 17.4 ms Why it matters: StartLux says general frameworks like vLLM are designed for token-by-token text generation and do not fit decision models, so it rewrote the engine around that workload. The company frames inference engineering as a growing competitive layer as model differences narrow. Coverage: 3 outlets in 24h

StartLux open-sourced Prism, described as the first inference engine built specifically for decision models, two days after DeepSeek open-sourced its DeepInfer framework.

StartLux says general frameworks like vLLM are designed for token-by-token text generation and do not fit decision models, so it rewrote the engine around that workload. The company frames inference engineering as a growing competitive layer as model differences narrow.