Skip to content

supported_models.md

Latest commit

 

History

History
55 lines (42 loc) · 2.94 KB

supported_models.md

File metadata and controls

55 lines (42 loc) · 2.94 KB
Raw

Supported Models

This catalog lists the maintained model-family guides published with TeleFuser. A model's Cookbook page is the source of truth for checkpoint identifiers, Hugging Face and ModelScope locations, required files, supported tasks, hardware assumptions, and runnable commands.

"Supported" means that TeleFuser contains a maintained pipeline or example contract. It does not mean that every checkpoint, precision, attention backend, or GPU topology is interchangeable. Use the validated profile documented by the selected guide before changing optimization settings.

World Models and Real-Time

Model Tasks Execution path
LingBot-World v2 Bidirectional world-model streaming Standalone and LiveKit
LingBot-World-Fast Causal interactive streaming Standalone and LiveKit
ABot-World-0-5B-LF Interactive generation HTTP and LiveKit

Video, Audio, and Restoration

Model Tasks Execution path
WanVideo T2V, I2V, FL2V Standalone and batch service
LTX-2.3 I2V with audio Standalone and batch service
LTX-2.5 Distilled T2V and I2V with audio Standalone and multi-GPU
MiniMax H3 T2VA, FL2VA, Ref2VA Standalone and batch service
LingBot-Video T2I, T2V, TI2V, refinement Standalone and multi-GPU
LongCat-Video T2V, I2V, continuation Standalone and batch service
LiveAct Speech-to-video Standalone and streaming
FlashVSR Video super-resolution Standalone and streaming
SwiftVR Causal video restoration Standalone and multi-GPU

Image Generation

Model Tasks Execution path
Qwen-Image T2I and editing Standalone and batch service
Z-Image T2I Standalone and batch service
Flux2 Klein T2I Standalone and batch service

Vision-Language-Action

Model Tasks Execution path
LingBot-VLA v2 Robot action prediction Standalone inference

Selecting a Profile

  1. Select a model family by task.
  2. Open its Cookbook page and use one of the documented model sources.
  3. Start from the validated hardware and precision profile.
  4. Run the documented standalone command before enabling serving or distributed execution.
  5. Change one optimization axis at a time and retain the guide's output validation.

For the smallest onboarding path, use the Wan2.1 1.3B profile in Basic Inference.