Gemma-4-26B-A4B-Maritime-zhTW-GGUF

云碩科技 · xCloudinfo · 系列:海事航海專業 · Maritime

繁體中文(台灣)+英文海事航海專業多模態模型的GGUF(llama.cpp / Ollama)。基於自家 Gemma-4-26B-A4B-TAIDE-zhTW 做雙語海事LoRA SFT,涵蓋避碰(COLREGs)、船舶操縱、導航、氣象、GMDSS、海事公約(SOLAS/MLC)、安全應急。含mmproj視覺檔,保留影像辨識。

做法

  • 資料(雙語海事SFT):英文海事知識整合自公開海事語料並自行蒸餾;繁體中文海事問答為云碩自行整理/編製的繁體中文海事語料(經自有流程蒸餾、清洗、繁體化);專有名詞(COLREGS/AIS/SOLAS/GMDSS)保留英文。
  • 訓練:於 云碩AI算力資源池(云碩算力中心) 以bf16 LoRA SFT → 合併 → GGUF。資料留在自有算力環境、流程可重現。

用法(llama.cpp,文字+視覺)

hf download xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW-GGUF --include "*Q4_K_M*.gguf" "mmproj-*.gguf" --local-dir .
# 文字+影像(給足回覆長度;Gemma-4 是 thinking 模型)
llama-server -m Gemma-4-26B-A4B-Maritime-zhTW-Q4_K_M.gguf \
  --mmproj mmproj-Gemma-4-26B-A4B-Maritime-zhTW-f16.gguf \
  -ngl 99 -c 4096 --host 0.0.0.0 --port 8080

Gemma-4為thinking模型,呼叫請給足 max_tokens(≥512),否則可能卡在推理、content空白。

用途與限制

  • 適用:海事教育、航海知識問答、COLREGs/公約解說。
  • 非航行決策依據 —— 實際操船避碰請依現行法規、官方海圖與船長判斷。事實查詢建議搭配RAG。

授權與來源聲明

  • 基底:xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW(gemma-4 = Apache-2.0;含TAIDE衍生,須遵守 TAIDE模型授權條款 並保留本聲明)。
  • 海事語料:繁中為云碩自行整理/編製;英文整合自公開海事語料並自行蒸餾。再散布/改作請保留本聲明並自行確認資料合規。
  • 不得用於軍事或非法用途;須遵守中華民國法律與適用之EU AI Act。

關於云碩科技

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。

由 云碩科技xCloudinfo於自有AI算力資源池製作;資料留在本地、流程可重現。

其他模型請見 https://hf.co/xCloudinfo。


Gemma-4-26B-A4B-Maritime-zhTW-GGUF — English

xCloudinfo (云碩科技) · Series: Maritime and navigation · Maritime

The GGUF (llama.cpp / Ollama) build of a Traditional Chinese (Taiwan) + English maritime and navigation multimodal model. It was built with bilingual maritime LoRA SFT on our own Gemma-4-26B-A4B-TAIDE-zhTW and covers collision avoidance (COLREGs), ship handling, navigation, meteorology, GMDSS, maritime conventions (SOLAS/MLC), and safety and emergency response. The mmproj vision file is included, so image recognition is retained.

Method

  • Data (bilingual maritime SFT): English maritime knowledge was integrated from public maritime corpora and distilled in-house; the Traditional Chinese maritime Q&A is a Traditional Chinese maritime corpus compiled and organized by xCloudinfo itself (distilled, cleaned, and converted to Traditional Chinese through our own pipeline); technical terms (COLREGS/AIS/SOLAS/GMDSS) are kept in English.
  • Training: bf16 LoRA SFT → merge → GGUF on the xCloudinfo AI compute resource pool (xCloudinfo compute center). The data stays in the in-house compute environment, and the process is reproducible.

Usage (llama.cpp, text + vision)

hf download xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW-GGUF --include "*Q4_K_M*.gguf" "mmproj-*.gguf" --local-dir .
# 文字+影像(給足回覆長度;Gemma-4 是 thinking 模型)
llama-server -m Gemma-4-26B-A4B-Maritime-zhTW-Q4_K_M.gguf \
  --mmproj mmproj-Gemma-4-26B-A4B-Maritime-zhTW-f16.gguf \
  -ngl 99 -c 4096 --host 0.0.0.0 --port 8080

Gemma-4 is a thinking model. When calling it, allow enough max_tokens (≥512); otherwise it may get stuck in reasoning and return empty content.

Intended use and limitations

  • Suitable for: maritime education, nautical knowledge Q&A, and explanations of COLREGs / conventions.
  • Not a basis for navigation decisions: for actual ship handling and collision avoidance, follow current regulations, official nautical charts, and the master's judgment. For factual queries, pairing it with RAG is recommended.

License and source statement

  • Base: xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW (gemma-4 = Apache-2.0; it includes TAIDE-derived components, so the TAIDE model license terms must be followed and this notice retained).
  • Maritime corpus: the Traditional Chinese part was compiled and organized by xCloudinfo; the English part was integrated from public maritime corpora and distilled in-house. When redistributing or adapting, retain this notice and confirm data compliance yourself.
  • Must not be used for military or illegal purposes; must comply with the laws of the Republic of China (Taiwan) and the applicable EU AI Act.

About xCloudinfo

xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.

Made by xCloudinfo (云碩科技) on its own AI compute resource pool; the data stays local and the process is reproducible.

More models: https://hf.co/xCloudinfo

Downloads last month
147
GGUF
Model size
25B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW-GGUF

Collection including xCloudinfo/Gemma-4-26B-A4B-Maritime-zhTW-GGUF