City96's Quantized UMT5-XXL Shrinks Wan Video Pipelines by 7GB City96's GGUF-quantized build of Google's UMT5-XXL text encoder has become the default encoder for local Wan 2.1/2.2 video pipelines, shrinking the 6-billion-parameter model from 22.7 GB in F32 to as little as 2.86 GB (Q3_K_S), with a Q4 file at about 3.5 GB. The Hugging Face repository reported more than 819,000 downloads in the cited month, and the build works with ComfyUI-GGUF, llama.cpp, Ollama, LM Studio and Docker Model Runner. City96 recommends Q5_K_M (4.15 GB) or larger because llama.cpp cannot build imatrix calibration for T5. City96's Quantized UMT5-XXL Shrinks Wan Video Pipelines by 7GB A GGUF-quantized build of Google's UMT5-XXL encoder is quietly powering most local Wan video pipelines, shrinking a 22GB model to under 4GB. - GGUF-quantized build of Google's UMT5-XXL encoder https://huggingface.co/google/umt5-xxl?utm source=alphasignal , 6B params, Apache 2.0 - Over 819,000 downloads last month, now the default text encoder for Wan video workflows - Quants range from 2.86 GB Q3 K S to 11.4 GB F16 , down from 22.7 GB in F32 - Q5 K M or larger recommended because llama.cpp cannot build imatrix calibration for T5 - Works with ComfyUI-GGUF https://github.com/city96/ComfyUI-GGUF?utm source=alphasignal , llama.cpp, Ollama, LM Studio and Docker Model Runner - Primary use case: Wan 2.1/2.2 text-to-video pipelines where VRAM budget is tight City96’s GGUF conversion https://huggingface.co/city96/umt5-xxl-encoder-gguf?utm source=alphasignal of Google’s UMT5-XXL encoder has become a common component in ComfyUI video workflows. The repository reported more than 819,000 downloads during the cited month. Hugging Face download counts represent file requests and can include repeat downloads, but the volume shows broad adoption across the open video stack. City96 repackaged the roughly six-billion-parameter encoder in quantized GGUF files that require far less memory than the original weights. A Q4 file occupies about 3.5 GB, leaving substantially more room for a 14B-class Wan diffusion transformer, VAE, runtime buffers, and optional image or audio components. One encoder, ten memory budgets The source checkpoint, google/umt5-xxl , is a multilingual T5 model. Video pipelines use its encoder to transform prompt tokens into conditioning vectors, so this repository omits the decoder and packages the required weights in GGUF, a format designed for efficient loading and quantized tensors. | Published GGUF files and approximate sizes | | | |---|---|---| | Quantization | Size | Practical use | |---|---|---| | Q3 K S | 2.86 GB | Severe memory constraints | | Q3 K M | 3.06 GB | Severe memory constraints | | Q4 K S | 3.50 GB | Low-VRAM workflows | | Q4 K M | 3.66 GB | Low-VRAM workflows | | Q5 K S | 4.05 GB | Balanced default | | Q5 K M | 4.15 GB | Publisher-recommended minimum | | Q6 K | 4.67 GB | Higher precision | | Q8 0 | 6.04 GB | Near-full precision | | F16 | 11.4 GB | Unquantized half precision | | F32 | 22.7 GB | Full precision | Compared with the 11.4 GB F16 file, This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.