Instructions to use GGQuanta/Xenomi-mini with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use GGQuanta/Xenomi-mini with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf GGQuanta/Xenomi-mini:BF16 # Run inference directly in the terminal: llama cli -hf GGQuanta/Xenomi-mini:BF16
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf GGQuanta/Xenomi-mini:BF16 # Run inference directly in the terminal: llama cli -hf GGQuanta/Xenomi-mini:BF16
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf GGQuanta/Xenomi-mini:BF16 # Run inference directly in the terminal: ./llama-cli -hf GGQuanta/Xenomi-mini:BF16
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf GGQuanta/Xenomi-mini:BF16 # Run inference directly in the terminal: ./build/bin/llama-cli -hf GGQuanta/Xenomi-mini:BF16
Use Docker
docker model run hf.co/GGQuanta/Xenomi-mini:BF16
- LM Studio
- Jan
- vLLM
How to use GGQuanta/Xenomi-mini with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "GGQuanta/Xenomi-mini" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "GGQuanta/Xenomi-mini", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/GGQuanta/Xenomi-mini:BF16
- Ollama
How to use GGQuanta/Xenomi-mini with Ollama:
ollama run hf.co/GGQuanta/Xenomi-mini:BF16
- Unsloth Desktop
- Pi
How to use GGQuanta/Xenomi-mini with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf GGQuanta/Xenomi-mini:BF16
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "GGQuanta/Xenomi-mini:BF16" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use GGQuanta/Xenomi-mini with Docker Model Runner:
docker model run hf.co/GGQuanta/Xenomi-mini:BF16
- Lemonade
How to use GGQuanta/Xenomi-mini with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull GGQuanta/Xenomi-mini:BF16
Run and chat with the model
lemonade run user.Xenomi-mini-BF16
List all available models
lemonade list
- Hermes Agent
How to use GGQuanta/Xenomi-mini with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf GGQuanta/Xenomi-mini:BF16
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default GGQuanta/Xenomi-mini:BF16
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use GGQuanta/Xenomi-mini with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf GGQuanta/Xenomi-mini:BF16
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "GGQuanta/Xenomi-mini:BF16" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
| license: apache-2.0 | |
| language: | |
| - zh | |
| - en | |
| base_model: Qwen/Qwen3.5-4B | |
| base_model_relation: finetune | |
| library_name: gguf | |
| pipeline_tag: text-generation | |
| tags: | |
| - gguf | |
| - llama.cpp | |
| - qwen | |
| - qwen3.5 | |
| - xenomi | |
| - quantum-research | |
| model_name: Xenomi-mini | |
| # 玄幂 Xenomi-mini(4B)GGUF | |
| Xenomi-mini 是玄幂家族的桌面档,面向**对话、科研摘要 / 科普、内容风险与产品判断**。 | |
| 本仓库提供合并后的 GGUF,可用 llama.cpp / `llama-cli` / `llama-server` 直接加载。 | |
| - **出品:** 北京中科国光量子科技有限公司(国光量子 / GGQUANTA) | |
| - **基座:** [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B)(Apache 2.0) | |
| - **训练:** 垂类数据 + LoRA / SFT,再 merge 为完整权重并转为 GGUF | |
| - **角色:** 单卡桌面生成,不是端侧 nano,也不是机房旗舰档 | |
| 同家族:0.8B = [Xenomi-nano](https://huggingface.co/GGQuanta/Xenomi-nano),27B = Xenomi-pro(规划),122B-A10B = Xenomi-max(规划)。 | |
| ## 选哪个文件 | |
| | 场景 | 建议 | | |
| | --- | --- | | |
| | 本机先跑起来 / 内存有限 | **Q4_K_M**(约 2.6 GB) | | |
| | 产品对话、路由、NER、短文本 | **产品轨 BF16**(约 7.8 GB) | | |
| | 摘要、科普、格式、内容风险、Agent 决策 | **科研轨 BF16**(约 7.8 GB) | | |
| | 文件 | 用途 | 体积 | | |
| | --- | --- | --- | | |
| | `xenomi-4b-sft-v0-id-r5-Q4_K_M.gguf` | 量化档,便于本机与端侧试用 | 约 2.6 GB | | |
| | `xenomi-4b-classic-v0-id8b-BF16.gguf` | 产品轨完整精度 | 约 7.8 GB | | |
| | `xenomi-4b-research-qrouter-v1-BF16.gguf` | 科研轨完整精度 | 约 7.8 GB | | |
| GGUF **不含**独立量子读出头。身份金句门(thin-A)是**服务侧规则**,权重本身不内置拦截。 | |
| 量化档便于部署,**不能**把产品 / 科研 BF16 的五维或 Layer A 分数直接写成该 Q4 文件的实测结果。 | |
| ## 重要声明 | |
| - **相对基座的准确率优势**归因于垂类数据与 LoRA / SFT,**不**归因于量子硬件或「量子优越性」。 | |
| - 产品五维、科研 Layer A 来自训练侧 **CUDA LoRA(PEFT)** 对照(RTX 5070,bf16,约 2026-07-31 至 2026-08-06)。对应 BF16 GGUF 为 adapter 合并转出,**未在 llama.cpp 上重跑同一套 eval**。 | |
| - Agent 决策三项来自 **llama.cpp BF16** 与当场转出的 Qwen3.5-4B 基座同口径(2026-08-18),**不与 CUDA PEFT 历史分混写**。 | |
| - Acc / F1 按任务分别呈现,不能加总为「通用能力分」。部分任务端到端更快,主要来自更短、更规范的生成;**不宣称 tokens/s 提升**。 | |
| - 已知回退:产品轨 RAG 87.5 → 85.0;科研轨摘要 must 50.0 → 45.8。 | |
| ## 评测 | |
| ### 产品轨(相对 Qwen3.5-4B) | |
| | 模型 | 路由 Acc | RAG Acc | NER micro-F1 | Keyword F1 | 短文本 Acc | Parse-OK | | |
| | --- | ---: | ---: | ---: | ---: | ---: | ---: | | |
| | Qwen3.5-4B | 70.0 | 87.5 | 50.4 | 21.8 | 76.2 | 100.0 | | |
| | Xenomi-mini 产品 | **95.0** | 85.0 | **78.1** | **60.5** | **100.0** | 100.0 | | |
| Identity FAQ(`--no-thin-a`):**91%**。 | |
| ### 科研轨 Layer A(相对 Qwen3.5-4B) | |
| | 模型 | 摘要槽位 | 摘要 must | 科普 must | Format | 风险 F1 | | |
| | --- | ---: | ---: | ---: | ---: | ---: | | |
| | Qwen3.5-4B | 100.0 | 50.0 | 14.6 | 95.8 | 44.2 | | |
| | Xenomi-mini 科研 | 100.0 | 45.8 | **41.7** | **100.0** | **100.0** | | |
| ### Agent 决策(llama.cpp BF16,相对同口径基座) | |
| | 任务 | 基座 | Xenomi-mini 科研 | | |
| | --- | ---: | ---: | | |
| | 风险决策 | 57.4 | **95.9** | | |
| | 工具决策 | 92.5 | **95.4** | | |
| | 结束决策 | 33.3 | **70.8** | | |
| ## 用法 | |
| ```bash | |
| # 量化档(推荐先试这个) | |
| llama-cli -hf GGQuanta/Xenomi-mini:Q4_K_M -cnv --jinja --reasoning off -c 8192 | |
| # 产品轨 / 科研轨 BF16:先下载对应文件 | |
| hf download GGQuanta/Xenomi-mini xenomi-4b-classic-v0-id8b-BF16.gguf | |
| llama-cli -m xenomi-4b-classic-v0-id8b-BF16.gguf -cnv --jinja --reasoning off -c 8192 | |
| hf download GGQuanta/Xenomi-mini xenomi-4b-research-qrouter-v1-BF16.gguf | |
| llama-cli -m xenomi-4b-research-qrouter-v1-BF16.gguf -cnv --jinja --reasoning off -c 8192 | |
| # OpenAI 兼容服务 | |
| llama-server -m xenomi-4b-sft-v0-id-r5-Q4_K_M.gguf --jinja --port 8080 | |
| ``` | |
| 仓库里有两份 BF16,`-hf …:BF16` 会有歧义,请按文件名下载。Chat 模板为 Qwen3.5 / ChatML。默认可关思考(`--reasoning off`)。 | |
| 建议系统提示(可选): | |
| ```text | |
| 你是玄幂(Xenomi),由中科国光量超训练的量子计算 / 科研领域模型。 | |
| ``` | |
| 非思考采样可从 `temperature=0.7`、`top_p=0.8`、`top_k=20` 起步。 | |
| ## 许可 | |
| 基座与本仓库权重按 **Apache 2.0** 提供。使用时请同时遵守 Qwen3.5 的许可与使用政策。 | |