Downloads · 30 days
1.4K
24% of all-time downloads
xCloudinfo/Gemma-4-31B-xVITA-zhTW-GGUF
Gemma-4-31B-xVITA-zhTW-GGUF is a image-text-to-text model from xCloudinfo. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Downloads · 30 days
1.4K
24% of all-time downloads
All-time downloads
5.9K
Public
Repo size
127 GB
Likes
0
Public
Click a slice to open those files.
.gguf127 GB · 100%
From the Hugging Face model README
云碩科技 · xCloudinfo · 系列:xVITA · 台灣通用繁中
以 google/gemma-4-31B-it(dense、31B、多模態指令模型)為基底,使用云碩自有「台灣通用繁體中文」語料(v5)微調而成的 GGUF(llama.cpp / Ollama) 版本,含視覺投影器(mmproj),可圖文對話。語料以真實瀏覽器開網頁抓取原文改寫、標註出處,未經第三方模型蒸餾。六階量化齊全,依裝置VRAM/記憶體選用。
本輪微調僅針對語言端投影層,視覺塔完全未更動,沿用母模原版;已實測驗證(真實圖片辨識形狀、顏色、文字皆正確)。母模同時具備音訊塔,但本次未轉出音訊投影器,僅提供視覺。
| 檔案 | 量化 | 大小 | 用途 |
|---|---|---|---|
model-Q8_0.gguf | Q8_0 | 32.6 GB | 幾乎無損,資源充足時首選 |
model-Q6_K.gguf | Q6_K | 25.2 GB | 高品質,效能/精度平衡 |
model-Q5_K_M.gguf | Q5_K_M | 21.8 GB | 品質與體積折衷 |
model-Q4_K_M.gguf | Q4_K_M | 18.7 GB | 部署首選,一般裝置最常用 |
model-IQ4_XS.gguf | IQ4_XS | 16.7 GB | 較小體積,帶imatrix校準 |
model-IQ2_M.gguf | IQ2_M | 10.9 GB | 極限壓縮,記憶體受限裝置 |
imatrix.dat | — | 13.8 MB | 量化用重要性矩陣(IQ系列必須) |
mmproj-Gemma-4-31B-xVITA-zhTW-f16.gguf | f16 | 1.2 GB | 視覺投影器(vision) |
llama-server -m model-Q4_K_M.gguf \
--mmproj mmproj-Gemma-4-31B-xVITA-zhTW-f16.gguf -c 4096 -ngl 99
純文字可省略
--mmproj。把model-Q4_K_M.gguf換成其他階的檔名即可切換量化等級。
google/gemma-4-31B-it,Apache-2.0 + Google Gemma 4 License(含Prohibited Use Policy)。由 云碩科技xCloudinfo於自有AI算力資源池製作。
云碩科技(xCloudinfo Corp. Limited)專注於地端私有化人工智慧解決方案。xVITA是云碩自研模型的品牌名稱。
其他模型請見 https://hf.co/xCloudinfo。
xCloudinfo (云碩科技) · Series: xVITA · Taiwan general-purpose Traditional Chinese
A GGUF (llama.cpp / Ollama) version built on google/gemma-4-31B-it (a dense, 31B, multimodal instruction model), fine-tuned with xCloudinfo's own "Taiwan general-purpose Traditional Chinese" corpus (v5), including the vision projector (mmproj), so it supports image-text conversation. The corpus was harvested by opening web pages in a real browser, rewriting the source text and annotating provenance, without distillation from any third-party model. All six quantization tiers are available; choose by your device's VRAM/memory.
This round of fine-tuning targets only the language-side projection layers; the vision tower is completely untouched and reuses the parent model's original version. This has been measured and verified (shapes, colors, and text in real images are all recognized correctly). The parent model also has an audio tower, but this release did not convert the audio projector and provides vision only.
| File | Quant | Size | Use |
|---|---|---|---|
model-Q8_0.gguf | Q8_0 | 32.6 GB | Near-lossless, first choice when resources are ample |
model-Q6_K.gguf | Q6_K | 25.2 GB | High quality, performance/precision balance |
model-Q5_K_M.gguf | Q5_K_M | 21.8 GB | Quality/size compromise |
model-Q4_K_M.gguf | Q4_K_M | 18.7 GB | Deployment first choice, most common on general devices |
model-IQ4_XS.gguf | IQ4_XS | 16.7 GB | Smaller size, with imatrix calibration |
model-IQ2_M.gguf | IQ2_M | 10.9 GB | Extreme compression, memory-constrained devices |
imatrix.dat | — | 13.8 MB | Importance matrix for quantization (required by the IQ series) |
mmproj-Gemma-4-31B-xVITA-zhTW-f16.gguf | f16 | 1.2 GB | Vision projector (vision) |
llama-server -m model-Q4_K_M.gguf \
--mmproj mmproj-Gemma-4-31B-xVITA-zhTW-f16.gguf -c 4096 -ngl 99
For text only, you can omit
--mmproj. Swapmodel-Q4_K_M.gguffor another tier's filename to switch quantization level.
google/gemma-4-31B-it, Apache-2.0 + Google Gemma 4 License (including the Prohibited Use Policy).Made by xCloudinfo (云碩科技) on its own AI compute pool.
xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.
More models: https://hf.co/xCloudinfo