Downloads · 30 days
330
26% of all-time downloads
xCloudinfo/Nemotron-3-Nano-30B-TAIDE-zhTW-Uncensored-GGUF
Nemotron-3-Nano-30B-TAIDE-zhTW-Uncensored-GGUF is a machine learning model from xCloudinfo. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Downloads · 30 days
330
26% of all-time downloads
All-time downloads
1.3K
Public
Repo size
155 GB
Likes
0
Public
Click a slice to open those files.
.gguf155 GB · 100%
From the Hugging Face model README
云碩科技 · xCloudinfo
以 nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16(nemotron_h_moe混合架構:Mamba-2 + Attention + latent-MoE,約31.6B總參 / 3B活躍)為底座,疊上繁體中文(台灣)與無審查合規的微調,保留底座原有的強程式能力。
功能:以道地台灣繁中作答的通用助理,程式能力幾乎完整保留底座水準,並針對資安教育等正當用途減少不必要的拒答。
許多在地化/客製微調為了塞進新語言或新行為,會默默把底座原有的程式與推理能力洗壞,卻不對外揭露。本模型反其道而行:加重程式資料replay,並以標準benchmark對照底座公開驗證。
| 項目 | 本模型 | 原始底座 | 差距 |
|---|---|---|---|
| 程式能力HumanEval pass@1(164題) | 87.2% | 90.2% | −3.0(幾乎零損失) |
| 繁中台灣知識MCQ | 86.4% | — | 在地化完整 |
| 作答誠實度(不確定時不硬掰) | 改善 | — | — |
| File | Quant | Size |
|---|---|---|
*-Q4_K_M.gguf | Q4_K_M | ~18 GB |
*-Q6_K.gguf | Q6_K | ~26 GB |
*-Q8_0.gguf | Q8_0 | ~33 GB |
*-f16.gguf | F16 | ~63 GB |
本模型架構為 nemotron_h_moe,需使用近期版本的 llama.cpp(已內建nemotron_h_moe支援)。舊版不認此架構,請自行編譯近期版:
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j # 純 CPU 可省略 -DGGML_CUDA=ON
啟動OpenAI相容服務:
llama-server -m Nemotron-3-Nano-30B-TAIDE-zhTW-Uncensored-Q6_K.gguf -ngl 99 -c 32768 --host 0.0.0.0 --port 8080
架構
nemotron_h_moe(混合Mamba-2 + Attention + latent-MoE);MTP(multi-token prediction)張量保留但llama.cpp尚未使用。
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16,NVIDIA Open Model License(衍生需保留來源通知)。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。
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)
Built on nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 (the nemotron_h_moe hybrid architecture: Mamba-2 + Attention + latent-MoE, about 31.6B total params / 3B active), with a Traditional Chinese (Taiwan) and uncensored-compliance fine-tune layered on top, preserving the base model's strong coding ability.
Function: a general-purpose assistant that answers in idiomatic Taiwan Traditional Chinese, with coding ability almost fully preserved at the base-model level, and fewer unnecessary refusals for legitimate uses such as security education.
Many localization/customization fine-tunes, in order to cram in a new language or new behavior, quietly ruin the base model's original coding and reasoning ability without disclosing it. This model does the opposite: we increase the replay of coding data and verify it openly against standard benchmarks relative to the base model.
| Item | This model | Original base | Gap |
|---|---|---|---|
| Coding ability, HumanEval pass@1 (164 problems) | 87.2% | 90.2% | −3.0 (almost no loss) |
| Traditional Chinese Taiwan knowledge MCQ | 86.4% | — | Localization is complete |
| Answer honesty (not making things up when uncertain) | Improved | — | — |
| File | Quant | Size |
|---|---|---|
*-Q4_K_M.gguf | Q4_K_M | ~18 GB |
*-Q6_K.gguf | Q6_K | ~26 GB |
*-Q8_0.gguf | Q8_0 | ~33 GB |
*-f16.gguf | F16 | ~63 GB |
This model's architecture is nemotron_h_moe and requires a recent version of llama.cpp (which has built-in nemotron_h_moe support). Older versions do not recognize this architecture, so build a recent version yourself:
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j # for CPU only, omit -DGGML_CUDA=ON
Start an OpenAI-compatible service:
llama-server -m Nemotron-3-Nano-30B-TAIDE-zhTW-Uncensored-Q6_K.gguf -ngl 99 -c 32768 --host 0.0.0.0 --port 8080
Architecture
nemotron_h_moe(hybrid Mamba-2 + Attention + latent-MoE); the MTP (multi-token prediction) tensors are retained but not yet used by llama.cpp.
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16, NVIDIA Open Model License (derivatives must retain the source notice).Made by xCloudinfo (云碩科技) on its own AI compute pool; data stays local, the process is reproducible.
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