LingBot-30B-3B_GGUF_ComfyUI

LingBot-Video-MoE-30B-A3B — GGUF (8GB-VRAM ready)

<div align="center"> <video src="https://cdn-uploads.huggingface.co/production/uploads/68761990332d15464ccc8dee/LMOmadbPymaHITY4nvbcv.mp4" width="400" height="400" autoplay loop muted playsinline ></video> </div> <div align="center"> <video src="https://cdn-uploads.huggingface.co/production/uploads/68761990332d15464ccc8dee/lU2xwAF1Xu2EqU9te9DG0.mp4" width="400" height="400" autoplay loop muted playsinline ></video> </div>

WARNING: Requires encoder and vae from 1.3b repo: https://huggingface.co/realrebelai/LingBot_ComfyUI/tree/main

GGUF quantizations of robbyant/lingbot-video-moe-30b-a3b — a 128-expert Mixture-of-Experts text-to-video model — packaged to run on consumer GPUs via ComfyUI_Rebels_LingBot. Tested on an RTX 3070 8GB / 16GB RAM.

How a 30B runs on 8GB

MoE "A3B" means ~3B active params per token, but a diffusion forward touches essentially all experts each step — so the win isn't compute, it's that the weights never need to be resident. The node loader keeps the quantized bytes in system RAM (memmap-backed) and dequantizes on demand: regular linears per forward, and the 128-expert fused stacks one expert matrix at a time, only for routed experts. Peak VRAM stays within 8GB; system RAM (as page cache) is what carries the model.

Because of this, VRAM use is the same across all quant tiers — you pick a tier by how much RAM you have, not VRAM.

Files

DiT GGUF GOES IN DIFFUSION MODELS FOLDER!

The expert tensors are quantized (not passed through at F16) — the fused 3D stacks are split to per-expert 2D matrices during conversion so the quantizer compresses them properly.

Also required (from the base repo / companion 1.3B repo): LingBot_vae.safetensors, a Qwen3-VL text encoder, and the 30B transformer/config.json (placed in the node pack's model_assets/ as transformer_config_30b.json).

Usage

Load with the LingBot 30B MoE Loader (GGUF) node; the rest of the graph (Structured Prompt → Text Encode → Sampler → Video Combine) is identical to the 1.3B workflow.

The model requires structured JSON captions, not prose — use the Structured Prompt node and always set lighting, or output drifts dark. See the GitHub pack for the full workflow.

Expect the first chunk to be slow (weights streaming from disk, page cache cold); it speeds up as the cache warms.

License

Other — inherits the upstream LingBot-Video license from robbyant/lingbot-video-moe-30b-a3b; see that repo's LICENSE before redistribution or commercial use. Quantization is a format conversion only. Quants & nodes by RealRebelAI.