<h3 id="gguf-file:">GGUF FILE:</h3><h3 id="https:huggingface.corealrebelailingbot-30b-3b_gguf_comfyuitreemainwarning:-requires-encoder-and-vae-from-1.3b-repo:-https:huggingface.corealrebelailingbot_comfyuitreemain"><a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai/LingBot-30B-3B_GGUF_ComfyUI/tree/main">https://huggingface.co/realrebelai/LingBot-30B-3B_GGUF_ComfyUI/tree/main</a> WARNING: Requires encoder and vae from 1.3b repo: <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai/LingBot_ComfyUI/tree/main">https://huggingface.co/realrebelai/LingBot_ComfyUI/tree/main</a></h3><hr /><hr /><hr /><hr /><hr /><hr /><hr />GGUF quantizations of <a target="_blank" rel="ugc" href="https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b">robbyant/lingbot-video-moe-30b-a3b</a> — a 128-expert Mixture-of-Experts text-to-video model — packaged to run on consumer GPUs via <a target="_blank" rel="ugc" href="https://github.com/RealRebelAI/ComfyUI_Rebels_LingBot">ComfyUI_Rebels_LingBot</a>. Tested on an RTX 3070 8GB / 16GB RAM.<h2 id="how-a-30b-runs-on-8gb">How a 30B runs on 8GB</h2>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.<h2 id="files">Files</h2>FileSizeNotes<code>LingBot-Video-30B-A3B-Q3_K_M.gguf</code>~13 GBFits a 16GB machine's page cache → steps at RAM speed. Recommended for 16GB RAM.<code>LingBot-Video-30B-A3B-Q4_K_M.gguf</code>~17 GBBest quality/size for 32GB-RAM machines.higher tiers (Q5/Q6/Q8)20–32 GBOnly worthwhile with large RAM.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): <code>LingBot_vae.safetensors</code>, a Qwen3-VL text encoder, and the 30B <code>transformer/config.json</code> (placed in the node pack's <code>model_assets/</code> as <code>transformer_config_30b.json</code>).<h2 id="usage">Usage</h2>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.<h2 id="license">License</h2>Other — inherits the upstream LingBot-Video license from <a target="_blank" rel="ugc" href="https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b">robbyant/lingbot-video-moe-30b-a3b</a>; see that repo's LICENSE before redistribution or commercial use. Quantization is a format conversion only. Quants &amp; nodes by <a target="_blank" rel="ugc" href="https://huggingface.co/realrebelai">RealRebelAI</a>.

Files with SHA256: cb1bd34c752a2b715f292fee36fc0980c6426268566b25e134861aa8fc9fe8bf