←Story detail

Trending story

StartLux open-sources decision model inference engine Prism

2 reports1 discussions3 story updatesUpdated Oct 23 19:50

Understand this story

AI overview

On October 22, 2026, StartLux (Yuandian Xinghui) open-sourced its self-developed inference engine Prism, calling it the first inference engine built specifically for decision models and open-sourced to the public. According to Zhidongxi reports, on a single H200, Prism processes a single short question 2.8 times faster than the original vLLM solution, and when asking three questions in a row on the same material, it achieves a maximum speedup of 5.3 times. The team says the speedup comes from four changes: narrowing the output computation range, forest-style computation, int8 integer arithmetic, and mixed-batch scheduling; in a 231-question cross-engine comparison, the average per-question latency was 41 milliseconds, lower than SGLang's 75.6 milliseconds and vLLM's 112.7 milliseconds. The team also released the complete engine code and optimization logic. There are currently no new developments following this open-source release.

AI-generated from public reports

Latest development · Report

StartLux open-sources the Prism inference engine for decision models, delivering up to 5.3x faster single-card inference

Developments

1 confirmed developments

  1. StartLux open-sources decision model inference engine Prism

    Source:Zhidx

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Oct 23

  1. Zhidongxi

    StartLux open-sources the Prism inference engine for decision models, delivering up to 5.3x faster single-card inference

    StartLux (Yuandian Xinghui) has open-sourced its self-developed inference engine Prism, designed specifically for decision models. On a single H200, it processes a single short question 2.8x faster than the original vLLM solution, and when asking three questions in a row on the same material, it achieves up to 5.3x speedup. Prism restructures the inference pipeline around four changes: narrowing the output computation range, forest-style computation, int8 integer arithmetic, and mixed-batch scheduling. In a 231-question cross-engine comparison, its average per-question latency was 41 milliseconds, lower than SGLang's 75.6 milliseconds and vLLM's 112.7 milliseconds. The team also released the complete engine code and optimization logic. Its decision model StartLux-Decision-v1.1 previously topped the Hugging Face Decision Index 0.3.2 with a score of 69.1.

  2. Phoenix Technology

    StartLux open-sources decision model inference engine Prism, up to 5.3x faster inference on a single card

    StartLux Origin Star open-sources its self-developed inference engine Prism, claiming it is the first inference engine built specifically for decision models and open-sourced. On a single H200, Prism processes a single short question 2.8x faster than the original vLLM 0.31 solution, and asking three questions in a row about the same material is up to 5.3x faster. The team says it achieves the speedup through four optimizations: narrowing the output layer computation range, forest-style computation, int8 integer arithmetic, and mixed scheduling; in a 231-question cross-engine comparison, the average per-question latency is 41 milliseconds.

  3. Phoenix TechnologyRelated discussion

    Phoenix Technology joined this discussion

    This post contributes to heat; it is not a formal report or confirmed fact.

Heat trend for this story

Current
5
Period peak
14 Oct 24 09:00
24-hour change
—

Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.

Related stories

  1. Related storiesDeepSeek open-sources inference framework DeepInfer
  2. Related storiesSGLang natively supports and powers the Beam production API
  3. Related storiesDeepSeek releases DeepSeek-V4.1-Flash-1017
  4. Related storiesStepFun updates Step Code to v0.1.4