AI Trends

Trending story

inclusionAI open-sources the Invoker-Render inference codebase

1 reports1 story updatesUpdated Nov 17 18:47

Understand this story

AI overview

On November 17, 2026, Ant's inclusionAI released the inference codebase for Code World Model 2.0 and its companion renderer Invoker-Render on GitHub. The approach keeps the world state in the upstream engine and has the video model handle rendering, with the two connected via structured proxy videos of fixed camera paths, scene layouts, and entity motion. Invoker-Render receives structured proxies, a single appearance reference image, and text; it is built on MiniMax-H3, SGLang, and SGLang-Omni, generates each chunk in four steps, and supports streaming output of arbitrary duration. On 140 held-out rollouts, it achieves an overall Elo of 1343.2, higher than SolarRenderer (Causal)'s 1245.3, StreamRender-H3's 1051.9, and Streaming CWM's 359.6. The repository provides an eight-GPU WebSocket session service, a file client, and a browser client; the model weights are hosted on Hugging Face, and the technical report has not yet been released.

AI-generated from public reports

Latest development · Report

Ant inclusionAI Open-Sources Inference Code for Code World Model 2.0 and Invoker-Render

Developments

1 confirmed developments

  1. inclusionAI open-sources the Invoker-Render inference codebase

    Source:English text is being prepared.

Reports and discussion timeline

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

Nov 17

  1. Ant inclusionAI: New GitHub RepositoryFirst-party

    Ant inclusionAI Open-Sources Inference Code for Code World Model 2.0 and Invoker-Render

    Ant inclusionAI has released on GitHub the inference code repository for Code World Model 2.0 and its companion renderer Invoker-Render, keeping world state in the upstream engine while the video model handles rendering, with the two connected via structured proxy videos of fixed camera paths, scene layouts, and entity motion. Invoker-Render receives structured proxies, a single appearance reference image, and text; it is built on MiniMax-H3, SGLang, and SGLang-Omni, generates each chunk in four steps, and supports streaming output of arbitrary duration. On 140 held-out rollouts, its composite Elo is 1343.2, higher than SolarRenderer (Causal)'s 1245.3, StreamRender-H3's 1051.9, and Streaming CWM's 359.6. The repository provides an eight-GPU WebSocket session service, a file client, and a browser client; model weights are hosted on Hugging Face, and the technical report has not yet been released.

Related stories

  1. Related storiesEnglish translation is being prepared.
  2. Related storiesHao AI Lab integrates MiniMax H3 Flex into FastVideo
  3. Related storiesNVIDIA and MIT Team Release Open-H3-Regenerate-2K Model Weights
  4. Related storiesStepFun updates Step Code to v0.1.4