infosave / MiniCPM5-2B-cmf

MiniCPM5-2B — CMF

MiniCPM5-2B (2.5B parameters, 42 layers, 128k context, thinking mode, tool calling; English and Chinese) packaged as single CMF files. They run with cortiq , a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, Apple silicon through Metal, and the CPU.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/MiniCPM5-2B-cmf MiniCPM5-2B-q8_2f.cmf --local-dir .
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "Explain quicksort in three sentences." --no-think

Use cortiq 0.7.6 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page . The engine picks the GPU and its settings itself. --no-think asks for a direct answer; without it the model reasons in a <think> block first, so raise --max-tokens (default 256).

Files

Both files are quantized from the bf16 checkpoint; the 4-bit weights are rounded with GPTQ on 2k tokens of wikitext-2 train. Perplexity: twelve 512-token windows of wikitext-2 test; the unquantized model scores 17.81 on the same windows.

Performance

One stream, medians of three runs:

The chat and prompt columns are measured through cortiq serve after one warm-up request, with the recommended sampling below.

Usage

cortiq run MiniCPM5-2B-q8_2f.cmf                             # interactive chat
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..."              # one answer, with reasoning
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --no-think   # answer without the thinking block
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --greedy     # deterministic output

Sampling

OpenBMB recommends temperature 1.0, top-p 0.95, min-p 0; if the output starts repeating, add a repetition penalty of 1.05:

cortiq run MiniCPM5-2B-q8_2f.cmf --temperature 1.0 --top-p 0.95 --top-k 0 --min-p 0 --rep-penalty 1.0 --max-tokens 4096 --prompt "..."

Without flags the CLI uses temperature 0.7, top-p 0.9, top-k 40, min-p 0.05, repetition penalty 1.1 and at most 256 new tokens. --top-k 0 and --rep-penalty 1.0 switch those filters off. Answers stop at <|im_end|> or </s> .

OpenAI-compatible server

cortiq serve MiniCPM5-2B-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "MiniCPM5-2B-q8_2f", "temperature": 1.0, "top_p": 0.95,
       "messages": [{"role": "user", "content": "Hi!"}]}'

Endpoints: /v1/chat/completions , /v1/completions , /v1/models , plus a web dashboard on the same port. --host 127.0.0.1 keeps the server local. temperature: 0 means greedy decoding; "enable_thinking": false in the request gives a direct answer.

Tool calling

Send tools in the usual OpenAI format; the embedded template puts them in the system prompt the way MiniCPM5 was trained. The model calls a tool in XML:

<function name="get_weather"><param name="city">Paris</param></function>

From cortiq 0.7.6 the server turns this into message.tool_calls with finish_reason: "tool_calls" , streamed or not; arguments is a JSON string whose values follow the types in the tool's schema. Older versions leave the XML in message.content . Run the tool, then send the assistant turn back with tool_calls and the result as a {"role": "tool", ...} message. The call and the answer from the result were correct in 6 of 6 test conversations for both files, on Vulkan and on Apple silicon. The embedded template differs from the upstream file only in writing tojson without ensure_ascii=False , which gives the same output, and the four tool-markup tokens are marked as ordinary text. cortiq 0.7.6 and later also accept the upstream template and tokenizer unchanged.

Hardware

Apple silicon (Metal). Runs on the GPU out of the box.

Choosing a GPU (Vulkan). cortiq gpu lists the adapters; CMF_GPU_ADAPTER=1 (index or name substring) picks one. If the list is empty on Linux, install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 ; on headless machines set XDG_RUNTIME_DIR=/tmp .

Long context

The model supports 131 072 tokens; the KV cache holds 32 768 by default. Raise it with CMF_MAX_SEQ :

CMF_MAX_SEQ=131072 cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --max-tokens 4096

The cache takes about 86 KB per token: 2.8 GB at 32k tokens, 11.3 GB at the full 131k. Decoding slows down as the context grows (about 18 tok/s at 16k tokens for q4tp on the RTX 3090).

Verify the download

sha256sum -c MiniCPM5-2B-q8_2f.cmf.sha256
cortiq verify MiniCPM5-2B-q8_2f.cmf
cortiq info MiniCPM5-2B-q8_2f.cmf

Weights derive from OpenBMB's release and remain under its Apache-2.0 terms.

---

Документация на русском

MiniCPM5-2B (2.5 млрд параметров, 42 слоя, контекст 128k, режим рассуждений, вызов инструментов; английский и китайский) в виде отдельных файлов CMF . Запускается движком cortiq на Rust без Python и ML-фреймворков: GPU NVIDIA, AMD и Intel через Vulkan/DX12, Apple silicon через Metal, а также CPU.

Быстрый старт

cargo install cortiq-cli
hf download infosave/MiniCPM5-2B-cmf MiniCPM5-2B-q8_2f.cmf --local-dir .
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "Объясни квиксорт в трёх предложениях." --no-think

Используйте cortiq 0.7.6 или новее; готовые сборки для Linux, macOS и Windows — на странице релизов . Движок сам выбирает GPU и настройки. --no-think даёт прямой ответ; без него модель сначала рассуждает в блоке <think> , поэтому увеличьте --max-tokens (по умолчанию 256).

Файлы

Оба файла квантованы из bf16-чекпойнта; 4-битные веса округлены GPTQ на 2k токенах wikitext-2 train. Перплексия измерена на 12 окнах wikitext-2 test по 512 токенов; неквантованная модель на тех же окнах даёт 17.81.

Скорость

Один поток, медианы трёх запусков:

Столбцы «чат» и «промпт» измерены через cortiq serve после одного прогревочного запроса, с рекомендуемым сэмплированием (ниже).

Использование

cortiq run MiniCPM5-2B-q8_2f.cmf                             # интерактивный чат
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..."              # один ответ с рассуждением
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --no-think   # ответ без блока рассуждений
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --greedy     # детерминированный вывод

OpenBMB рекомендует temperature 1.0, top-p 0.95, min-p 0; если вывод начинает повторяться, добавьте repetition penalty 1.05:

cortiq run MiniCPM5-2B-q8_2f.cmf --temperature 1.0 --top-p 0.95 --top-k 0 --min-p 0 --rep-penalty 1.0 --max-tokens 4096 --prompt "..."

Без флагов CLI использует temperature 0.7, top-p 0.9, top-k 40, min-p 0.05, repetition penalty 1.1 и не более 256 новых токенов. --top-k 0 и --rep-penalty 1.0 отключают эти фильтры. Ответ заканчивается на <|im_end|> или </s> .

Сервер с OpenAI-совместимым API:

cortiq serve MiniCPM5-2B-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "MiniCPM5-2B-q8_2f", "temperature": 1.0, "top_p": 0.95,
       "messages": [{"role": "user", "content": "Привет!"}]}'

Эндпоинты /v1/chat/completions , /v1/completions , /v1/models и веб-панель на том же порту. --host 127.0.0.1 оставляет сервер локальным. temperature: 0 означает greedy-декодирование; "enable_thinking": false в запросе даёт прямой ответ.

Вызов инструментов. Передавайте tools в обычном формате OpenAI; встроенный шаблон помещает их в системный промпт так, как модель обучали. Модель вызывает инструмент в XML:

<function name="get_weather"><param name="city">Paris</param></function>

Начиная с cortiq 0.7.6 сервер превращает такой вызов в message.tool_calls с finish_reason: "tool_calls" , и в потоковом режиме тоже; arguments — строка JSON, значения в которой приведены к типам из схемы инструмента. Более старые версии оставляют XML в message.content . Выполните инструмент и отправьте обратно ход ассистента с tool_calls и результат сообщением {"role": "tool", ...} . Вызов и ответ по результату были верны в 6 из 6 тестовых диалогов для обоих файлов, на Vulkan и на Apple silicon. Встроенный шаблон отличается от исходного только тем, что tojson записан без ensure_ascii=False (результат тот же), а четыре токена разметки вызова помечены как обычный текст. cortiq 0.7.6 и новее принимает и исходные шаблон и токенизатор без изменений.

Железо

Apple silicon (Metal). Работает на GPU без настройки.

Выбор GPU (Vulkan). cortiq gpu показывает адаптеры; CMF_GPU_ADAPTER=1 (индекс или часть имени) закрепляет нужный. Если на Linux список пуст, установите libvulkan1 libglvnd0 libegl1 libgl1 libglx0 ; на машинах без дисплея задайте XDG_RUNTIME_DIR=/tmp .

Длинный контекст

Модель поддерживает 131 072 токена; KV-кэш по умолчанию рассчитан на 32 768. Увеличьте его через CMF_MAX_SEQ : CMF_MAX_SEQ=131072 cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --max-tokens 4096 .

Кэш занимает около 86 КБ на токен: 2.8 ГБ на 32k токенов, 11.3 ГБ на полных 131k. С ростом контекста декодирование замедляется (около 18 ток/с на 16k токенов для q4tp на RTX 3090).

Проверка загрузки

sha256sum -c MiniCPM5-2B-q8_2f.cmf.sha256
cortiq verify MiniCPM5-2B-q8_2f.cmf
cortiq info MiniCPM5-2B-q8_2f.cmf

Веса получены из релиза OpenBMB и распространяются на условиях Apache-2.0.

---

中文文档

MiniCPM5-2B （25 亿参数，42 层， 128k 上下文，支持思考模式和工具调用，中英双语）打包为独立的 CMF 文件，由 cortiq 运行——一个不依赖 Python 和机器学习框架的 Rust 推理引擎：NVIDIA、AMD、Intel 显卡走 Vulkan/DX12， Apple silicon 走 Metal，也可在 CPU 上运行。

快速开始

cargo install cortiq-cli
hf download infosave/MiniCPM5-2B-cmf MiniCPM5-2B-q8_2f.cmf --local-dir .
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "用三句话解释快速排序。" --no-think

建议使用 cortiq 0.7.6 或更新版本；Linux、macOS、Windows 预编译二进制见 发布页面 。引擎会自动选择 GPU 和 设置。 --no-think 直接给出回答；不加该参数时模型会先在 <think> 块中思考， 请调高 --max-tokens （默认 256）。

文件

两个文件均由 bf16 检查点量化；4 位权重使用 GPTQ 在 2k 个 wikitext-2 train token 上 取整。困惑度基于 wikitext-2 test 的 12 个 512 token 窗口；未量化模型在相同窗口上为 17.81。

性能

单流，三次运行取中位数：

「对话」和「提示词」两列通过 cortiq serve 测得（先发一次预热请求），使用下文 推荐的采样参数。

使用

cortiq run MiniCPM5-2B-q8_2f.cmf                             # 交互式对话
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..."              # 单次回答（含思考）
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --no-think   # 不输出思考块
cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --greedy     # 确定性输出

OpenBMB 推荐 temperature 1.0、top-p 0.95、min-p 0；如出现重复输出，可加 repetition penalty 1.05：

cortiq run MiniCPM5-2B-q8_2f.cmf --temperature 1.0 --top-p 0.95 --top-k 0 --min-p 0 --rep-penalty 1.0 --max-tokens 4096 --prompt "..."

不加参数时 CLI 使用 temperature 0.7、top-p 0.9、top-k 40、min-p 0.05、 repetition penalty 1.1，最多生成 256 个新 token。 --top-k 0 和 --rep-penalty 1.0 关闭这两项过滤。回答在 <|im_end|> 或 </s> 处结束。

OpenAI 兼容 API 服务器：

cortiq serve MiniCPM5-2B-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "MiniCPM5-2B-q8_2f", "temperature": 1.0, "top_p": 0.95,
       "messages": [{"role": "user", "content": "你好！"}]}'

提供 /v1/chat/completions 、 /v1/completions 、 /v1/models 端点，同一端口还有 网页控制台。 --host 127.0.0.1 仅限本机访问。请求中 temperature: 0 表示贪心 解码； "enable_thinking": false 直接给出回答。

工具调用。 按常规 OpenAI 格式传入 tools ；内置模板会按 MiniCPM5 的训练方式 把它们放进系统提示词。模型以 XML 形式调用工具：

<function name="get_weather"><param name="city">Paris</param></function>

从 cortiq 0.7.6 起，服务器会把它转换为 message.tool_calls ，并给出 finish_reason: "tool_calls" ，流式输出同样如此； arguments 是 JSON 字符串，其中的值 按工具 schema 声明的类型转换。更早的版本会把 XML 留在 message.content 中。 执行工具后，把带 tool_calls 的助手回合和 {"role": "tool", ...} 结果消息一并发回。在 Vulkan 和 Apple silicon 上，两个文件在 6 个测试对话中调用和基于结果的回答全部正确。内置模板与上游文件的唯一区别是 tojson 去掉了 ensure_ascii=False （输出相同），并且四个工具标记 token 被标为普通文本。 cortiq 0.7.6 及以后版本也可以直接使用未经修改的上游模板和分词器。

硬件

Apple silicon（Metal）。 开箱即用 GPU 加速。

选择显卡（Vulkan）。 cortiq gpu 列出所有适配器； CMF_GPU_ADAPTER=1 （索引或 名称片段）指定使用哪一块。若 Linux 上列表为空，请安装 libvulkan1 libglvnd0 libegl1 libgl1 libglx0 ；无显示器的机器需设置 XDG_RUNTIME_DIR=/tmp 。

长上下文

模型支持 131 072 个 token；KV 缓存默认容纳 32 768 个，可用 CMF_MAX_SEQ 调高： CMF_MAX_SEQ=131072 cortiq run MiniCPM5-2B-q8_2f.cmf --prompt "..." --max-tokens 4096 。

缓存每个 token 约占 86 KB：32k token 为 2.8 GB，完整 131k 为 11.3 GB。上下文越长 解码越慢（RTX 3090 上 q4tp 在 16k token 时约 18 tok/s）。

校验下载

sha256sum -c MiniCPM5-2B-q8_2f.cmf.sha256
cortiq verify MiniCPM5-2B-q8_2f.cmf
cortiq info MiniCPM5-2B-q8_2f.cmf

权重来自 OpenBMB 的发布，遵循其 Apache-2.0 许可。