APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

City96's Quantized UMT5-XXL Shrinks Wan Video Pipelines by 7GB

GGUF-quantized build of Google's UMT5-XXL encoder , 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

City96's Quantized UMT5-XXL Shrinks Wan Video Pipelines by 7GB
Source
AlphaSignal
Published
Author
AlphaSignal
Read
1 min read

Takeaways

  • GGUF-quantized build of Google's UMT5-XXL encoder , 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 , 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 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.

QuantizationSizePractical use
Q3_K_S2.86 GBSevere memory constraints
Q3_K_M3.06 GBSevere memory constraints
Q4_K_S3.50 GBLow-VRAM workflows
Q4_K_M3.66 GBLow-VRAM workflows
Q5_K_S4.05 GBBalanced default
Q5_K_M4.15 GBPublisher-recommended minimum
Q6_K4.67 GBHigher precision
Q8_06.04 GBNear-full precision
F1611.4 GBUnquantized half precision
F3222.7 GBFull precision

Compared with the 11.4 GB F16 file,

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01huggingface.co
  2. 02github.com
  3. 03huggingface.co
  4. 04huggingface.co
  5. 05huggingface.co