Datasheet for local AI49 models98 GPUsData read 2026-09-25
No adsNo tracking
Check my GPU →

Text encoders: why T5, UMT5 and Qwen-VL eat your memory

The text encoder is often bigger than people expect, sometimes bigger than the image model. What it does and how to keep it out of your way.

G.05ExplainerUpdated 2026-09-25

Most newer image and video models come in parts you download separately (older SD 1.5 and SDXL checkpoints bundle everything): the diffusion model (the part that makes the picture) and the text encoder (the part that reads your prompt). The text encoder is often surprisingly large — and in 2026 it keeps growing, because new models use full vision-language models to read prompts.

How big they are

Text encoderUsed byLargest fileSmallest common file
T5-XXLFLUX.1, Chroma, SD 3.5, HiDream (plus Llama 3.1 8B)9.8 GB 16-bit2.9 GB GGUF Q4_K_M
UMT5-XXLWan 2.1, Wan 2.2, Wan Animate11.4 GB 16-bit3.7 GB GGUF Q4_K_M
Qwen2.5-VL 7BQwen-Image, Qwen-Image-Edit, HunyuanImage 2.116.6 GB 16-bit9.4 GB FP8
Qwen3 4BZ-Image, FLUX.2 [klein] 4B8.0 GB 16-bit5.6 GB FP8
Qwen3-VL 4BKrea 28.9 GB 16-bit5.2 GB FP8
Qwen3-VL 8BQwen-Image 2.117.5 GB 16-bit9.4 GB INT8
Mistral Small 3 (24B)FLUX.2 [dev]35.6 GB 16-bit18.0 GB FP8
Qwen3-VL 32BMiniMax H351.5 GB 16-bit15.7 GB NVFP4

File sizes from Hugging Face (Comfy-Org, comfyanonymous and city96 repos), read 2026-09-25.

Look at Wan 2.1 1.3B: the video model is 2.8 GB, its text encoder 11.4 GB at 16-bit. Four times bigger than the model. MiniMax H3 is the extreme: a 51.5 GB text encoder.

Good news: it does not need to fit at the same time

The text encoder runs once, at the start, to turn your prompt into numbers. Then ComfyUI can move it out of VRAM before the diffusion model starts working. So for "will it fit?" the diffusion model is what counts. That is why the verdicts on this site are based on the diffusion model file, with the encoder handled separately on each page.

What to do with it

  • Use the FP8 (or INT8) encoder. For text encoders the quality difference is tiny.
  • On small GPUs, use a GGUF encoder (city96 has T5 and UMT5 in many sizes) with the GGUF CLIP loader.
  • If it still does not fit, run it on the CPU. The images do not change; prompt encoding gets slower — by seconds for T5-class encoders, by a minute or more for the 24–32B ones.
  • It still has to fit in system RAM. That is where most "my PC froze" stories come from.