Load LingBot Video Model (Yogurt LingBot Video)

LingBot-Video is Robbyant's open-weights entry in the 2026 video-model wave, and the first thing everyone said when it dropped was "give me ComfyUI nodes." This loader is that support - with a twist. The interesting flavor, the MoE 30B-A3B, ships roughly 60 GB of BF16 transformer weights, about three times what your 4090 can hold. The entire point of this node is that it still works anyway. It doesn't quantize, it doesn't shrink, it doesn't stream from the cloud. It block-swaps.

The loader builds the full Diffusers-style pipeline from ComfyUI/models/LingBotVideo/: the transformer, a Wan-style VAE (the community clocked that immediately - "same vae as wan"), a Qwen3-VL text encoder, processor, and a flow-matching scheduler. And it's aggressively local. No Hugging Face cache, no background downloads, no telemetry. download_model defaults to false, so a missing model errors out instead of quietly pulling sixty gigabytes you didn't ask for.

The part that matters: MoE block swap

With cpu_offload on, the first moe_gpu_blocks transformer blocks stay permanently resident on the GPU; the rest live in system RAM and stream up for each step, then drop back down. Routing, expert selection, attention, and the sampling math are untouched - the pack's tests claim pixel-identical output to the official reference whether swap is on or off. On a 4090, moe_gpu_blocks=12 (the default) sits at roughly 19.4 GB stable and turns a 40-step 832×480 T2I into about six and a half minutes. Zero is the conservative dial - everything swaps, lowest VRAM, visibly slower. 14–16 is the aggressive band where any other model sharing VRAM can tip you into OOM. And 48 means "keep the whole thing on the GPU," which is not happening on 24 GB.

The inputs you actually touch:

• model_name - dropdown of what's in models/LingBotVideo: Dense 1.3B or MoE 30B-A3B.
• mode - t2i, t2v, or ti2v. This is baked into the loaded pipeline and must match what the Generate node will do; change it and you reload.
• transformer_dtype - auto or bfloat16 are the sane choices.
• cpu_offload and moe_gpu_blocks - your VRAM dials. Dense ignores moe_gpu_blocks.
• download_model - leave it off unless you really mean it.

Output is a single pipeline handle of the custom type YOGURT_LINGBOT_VIDEO_PIPELINE; wire it straight into LingBot Video Generate.

Install

Clone into custom_nodes, then install the requirements into the Python environment ComfyUI actually runs on - that bit trips people up:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes-LingBotVideo

pip install -r custom_nodes/ComfyUI-YogurtNodes-LingBotVideo/requirements.txt

The deps are transformers>=4.57,<5, diffusers>=0.36, accelerate, peft, safetensors, scipy, json_repair, and huggingface_hub. Restart ComfyUI after installing. Models keep the official Diffusers folder structure under:

ComfyUI/models/LingBotVideo/Robbyant--lingbot-video-dense-1.3b/
ComfyUI/models/LingBotVideo/Robbyant--lingbot-video-moe-30b-a3b/

When it misbehaves

Empty model dropdown means the folder structure is incomplete (it needs model_index.json plus transformer, text_encoder, processor, vae, and scheduler subfolders) - then restart. First MoE load takes minutes because it reads ~60 GB from disk; later runs reuse the same handle and are much faster. If changing moe_gpu_blocks seems to do nothing, ComfyUI cached the loader - re-queue, or restart ComfyUI. And 64 GB of system RAM runs, but sits at the edge; the author recommends 96 GB+, and given the pagefile-paging failure mode, it's hard to argue.