UMT5-XXL Encoder · GSQ/RCO GGUF

Calibrated text encoder for the Wan video pipeline.

Русский. Текстовый энкодер UMT5-XXL для видео Wan. Смешанные кванты Q3_K, Q4_K и Q6_K под заданный бюджет весов.

Evidence / Проверка

Recorded projection error 60,640 → 37,930. The encoder GGUF is 3.62 GB.

Ошибка проекций 60 640 → 37 930 на записанной выборке. GGUF энкодера занимает 3,62 ГБ.

Projection error measures calibration fit on the recorded workload. Ошибка проекций характеризует подгонку на записанной выборке.

Calibration runs / Замеры калибровки

Files / Файлы

Companions / Компоненты
Wan2.2-TI2V-5B-Turbo-GSQ-RCO-GGUF

Sources

Weights and hardware

Format: Q3_K, Q4_K, Q6_K. Tensor layout: Q6_K × 31, F32 × 73, Q4_K × 136, Q3_K × 2.

Vulkan GSQ calibrated 168 matrix projections. RCO selected Q3_K, Q4_K and Q6_K tensor layouts within the checkpoint byte budget. Remaining tensors use FP32.

Target system: Windows, AMD Radeon RX 580 2048SP (Polaris, 8 GB) and Intel Xeon E5-2696 v3 (Haswell, AVX2). Use as the Wan text encoder with a compatible Wan diffusion checkpoint and VAE. Media runtime

SHA-256 cec23a999e52f7edc7db5e7816469d8034d9a4ce5ece144ee013ea78f7f2b602.

Downloads last month
185
GGUF
Model size
6B params
Architecture
t5encoder
Hardware compatibility
Log In to add your hardware

6-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TalasCloud/UMT5-XXL-Encoder-GSQ-RCO-GGUF

Base model

google/umt5-xxl
Quantized
(11)
this model

Collection including TalasCloud/UMT5-XXL-Encoder-GSQ-RCO-GGUF