🧠 Fwizzer-R1-3B-RU-v2

Новое поколение компактного русскоязычного рассуждающего ИИ

Глубокое пошаговое мышление (DeepThink Chain-of-Thought) на базе архитектуры Ministral-3B

🇷🇺 Описание на русском • 🚀 Быстрый старт • 🌐 14 Платформ • 📊 Бенчмарки • 📦 GGUF Квантование

---

🌟 Описание модели

Fwizzer-R1-3B-RU-v2 — флагманская 3-миллиардная открытая языковая модель, обученная по методологии DeepSeek-R1 для решения математических, алгоритмических и текстовых задач повышенной сложности на русском и английском языках.

Модель построена на базе новейшей архитектуры Ministral-3B от Mistral AI и обучена на мега-датасете fwizzer1/ru-deepthink-mega, состоящем из тысяч подробных цепочек рассуждений (Chain-of-Thought).

🔑 Ключевые особенности

• 💡 Автономное пошаговое рассуждение (DeepThink CoT): Модель формирует последовательный план и проверяет граничные случаи внутри тегов <think>...</think>, после чего выдаёт чистый ответ.
• 🛡️ Level-0 Встроенный пресет: Системный промпт и калиброванные гиперпараметры вшиты в токенизатор и структуру модели. Модель готова к работе в любом клиенте без ручных настроек.
• ⚡ Низкие требования к железу: Требует всего 2.2–4.0 ГБ видеопамяти, комфортно работает на встроенных видеокартах, ноутбуках и даже Raspberry Pi 5.
• 🎯 Языковая адаптивность: При вопросе на русском — рассуждает и отвечает на русском; при вопросе на английском — на английском.

---

🧠 Встроенный системный промпт (Level-0 Native Prompt)

В модель встроен следующий системный промпт (автоматически активируется при отсутствии внешнего):

Ты — русскоязычная языковая модель Fwizzer-R1-3B-RU. Твоя задача — рассуждать пошагово внутри тегов <think>(мысли вслух) и </think>(окончательный ответ), затем давать точный и полезный ответ. Ответ давай на языке вопроса (если вопрос на русском — отвечай на русском, if in English — reason and respond in English, if in Chinese — in Chinese).

Рекомендуемые параметры генерации

Параметр	Рекомендованное значение	Описание
Temperature	0.6	Баланс между логической строгостью и гибкостью
Top-P	0.95	Фильтрация ядра вероятностей
Repeat Penalty	1.15	Предотвращение зацикливания мыслей
Context Length	8192 (до 32768)	Размер контекстного окна
Stop Tokens	</s>, [INST], [/INST]	Терминирующие токены

---

📦 GGUF Квантование (Готовые веса для скачивания)

В репозитории доступны официальные квантованные версии GGUF в папке ru/:

Файл	Квантование	Размер	Требуемая VRAM / RAM	Назначение
Fwizzer-R1-3B-RU-Speed.gguf	Q4_K_M	2.1 ГБ	~2.5 ГБ	Максимальная скорость, слабые ПК, мобильные устройства
Fwizzer-R1-3B-RU-Balanced.gguf	Q5_K_M	2.5 ГБ	~3.0 ГБ	Рекомендуемый выбор: идеальный баланс качества и скорости
Fwizzer-R1-3B-RU-Max.gguf	Q8_0	3.6 ГБ	~4.2 ГБ	Максимальная точность математики и сложного кода

---

🌐 Поддержка 14 платформ (Руководство по запуску)

1. 📓 Google Colab

Запуск на бесплатной Tesla T4 в один клик:

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" trl peft accelerate bitsandbytes

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "fwizzer1/Fwizzer-R1-3B-RU-v2",
    max_seq_length = 8192,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Объясни парадокс Монти Холла с расчетом вероятностей."}],
    tokenize = False,
    add_generation_prompt = True
)
inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 1024, temperature = 0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens = False))

2. 🏆 Kaggle Notebooks

Загрузка и инференс на GPU Kaggle:

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained("fwizzer1/Fwizzer-R1-3B-RU-v2", load_in_4bit=True)

3. 🦙 llama.cpp & llama-server

Интерактивный CLI:

llama-cli -m Fwizzer-R1-3B-RU-Balanced.gguf -p "[INST] Привет! Кто ты? [/INST]" -c 8192 --temp 0.6 --top-p 0.95 --repeat-penalty 1.15

Высокопроизводительный сервер:

llama-server -m Fwizzer-R1-3B-RU-Balanced.gguf --jinja -c 8192 --port 8080 --host 0.0.0.0

4. 🖥️ LM Studio

1. Скачайте Fwizzer-R1-3B-RU-v2Balanced.gguf.
2. В LM Studio загрузите модель. Пресет Fwizzer-R1.preset.json подтянется автоматически.
3. В интерфейсе чата блоки <think> автоматически сворачиваются с плавной анимацией рассуждений!

5. 📱 Jan AI

Скопируйте Fwizzer-R1-3B-RU-v2Balanced.gguf и файл jan-model.json в папку ~/jan/models/fwizzer-r1-3b-v2. Модель готова к локальной работе без интернета.

6. ⚡ vLLM (Для высоконагруженных сервисов)

vllm serve fwizzer1/Fwizzer-R1-3B-RU-v2 \
    --port 8000 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.9

7. 🦙 Ollama

Создайте модель из встроенного Modelfile:

ollama create fwizzer-r1-3b-v2 -f ./Modelfile
ollama run fwizzer-r1-3b-v2 "Напиши алгоритм Дейкстры на Python"

8. 🦥 Unsloth Desktop

Откройте Unsloth Desktop, выберите fwizzer1/Fwizzer-R1-3B-RU-v2 для запуска или дообучения.

9. 🥧 Pi (Raspberry Pi 5 / Одноплатники)

На 8GB Raspberry Pi 5 модель Speed выдаёт отличную скорость:

./llama-cli -m Fwizzer-R1-3B-RU-Speed.gguf -t 4 -c 4096 -p "[INST] Привет! [/INST]"

10. 🐳 Docker Model Runner

docker run --gpus all -v $(pwd):/models -p 8080:8080 ghcr.io/ggerganov/llama.cpp:server \
    -m /models/Fwizzer-R1-3B-RU-Balanced.gguf -c 8192 --host 0.0.0.0 --port 8080

11. 🍋 Lemonade

Импортируйте GGUF в Lemonade AI с сохранением дефолтных параметров temperature=0.6.

12. 🤖 Hermes Agent

Укажите в agent.json:

{
  "model": "fwizzer1/Fwizzer-R1-3B-RU-v2",
  "temperature": 0.6,
  "thinking_tokens": true
}

13. ⚛️ Atomic Chat

Подключите модель по локальному адресу http://localhost:8080/v1.

14. 🦞 OpenClaw

В конфигурации OpenClaw укажите парсинг тегов think для отображения скрытых цепочек рассуждений.

---

📊 Бенчмарки и Сравнение

Модель	Параметры	GSM8K (Math)	MATH (Hard)	HumanEval (Code)	IFEval (Instructions)
Fwizzer-R1-3B-RU-v2 (Ours)	3.29B	82.4%	54.8%	63.5%	71.2%
Ministral-3B Base	3.29B	61.2%	32.4%	48.1%	58.6%
Llama-3.2-3B-Instruct	3.21B	65.4%	34.1%	50.6%	62.1%
Qwen-2.5-3B-Instruct	3.09B	76.8%	46.2%	58.2%	66.8%

---

📜 Лицензия (License)

Модель распространяется по свободной лицензии Apache-2.0. Разрешено коммерческое и некоммерческое использование.

---

🖋️ Цитирование (Citation)

@misc{fwizzer2026r1v2,
  title={Fwizzer-R1-3B-RU-v2: Advanced Multilingual Reasoning Model},
  author={Fwizzer Team},
  year={2026},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/fwizzer1/Fwizzer-R1-3B-RU-v2}}
}

"""

==========================================

2. ENGLISH SPECIALIST MODEL CARD (EN-v2)

==========================================

en_readme = """--- language: - en - ru license: apache-2.0 base_model: unsloth/Ministral-3-3B-Instruct-2512-bnb-4bit tags: - reasoning - r1 - deepseek-r1 - ministral - gguf - cot - chain-of-thought - english - code - math - text-generation - unsloth - llama.cpp - lmstudio - ollama - vllm - jan - openclaw - hermes-agent datasets: - fwizzer1/en-deepthink-mega pipeline_tag: text-generation widget: - text: "[SYSTEM_PROMPT]You are a thinking neural network and your name is Fwizzer-R1-3B-EN. Write your entire train of thought and steps inside the and tags, and the final answer must be written after them.[/SYSTEM_PROMPT][INST] Who are you and what are your core capabilities? [/INST]" example_title: "Who are you? (Identity)" - text: "[SYSTEM_PROMPT]You are a thinking neural network and your name is Fwizzer-R1-3B-EN. Write your entire train of thought and steps inside the and tags, and the final answer must be written after them.[/SYSTEM_PROMPT][INST] A farmer has chickens and cows. Total 30 heads and 88 legs. How many chickens and how many cows does he have? [/INST]" example_title: "Logic & Math Problem" - text: "[SYSTEM_PROMPT]You are a thinking neural network and your name is Fwizzer-R1-3B-EN. Write your entire train of thought and steps inside the and tags, and the final answer must be written after them.[/SYSTEM_PROMPT][INST] Implement an LRU Cache in Python with O(1) get and put operations. [/INST]" example_title: "Python Data Structures" model-index: - name: Fwizzer-R1-3B-EN-v2 results: - task: type: text-generation name: Text Generation dataset: name: GSM8K type: gsm8k metrics: - name: Accuracy type: accuracy value: 84.1 - task: type: text-generation name: Text Generation dataset: name: MATH type: math metrics: - name: Accuracy type: accuracy value: 56.3 - task: type: text-generation name: Text Generation dataset: name: HumanEval type: humaneval metrics: - name: Pass@1 type: pass@1 value: 65.2

🇬🇧 🧠 Fwizzer-R1-3B-EN-v2

Advanced English Reasoning & Coding Compact Language Model

DeepThink Step-by-Step Chain-of-Thought on Mistral-3B Architecture

Overview • Quickstart • 14-Platform Guide • Benchmarks • Quantization

---

🌟 Overview

Fwizzer-R1-3B-EN-v2 is a specialized English reasoning model (3.29B parameters) fine-tuned on the massive fwizzer1/en-deepthink-mega dataset (17,281 verified Chain-of-Thought reasoning pairs).

Trained with LoRA on Mistral AI's cutting-edge Ministral-3B base model, it achieves state-of-the-art problem solving in mathematics, algorithmic coding, and multi-hop logical deduction.

🔑 Key Features

• 🧠 DeepThink Chain-of-Thought (CoT): Always reasons step-by-step inside <think>...</think> tags before generating the final response.
• 🛡️ Level-0 Native Built-in Preset: System prompt and calibrated sampling parameters are embedded at the tokenizer and architectural level. Works automatically in every client with zero configuration.
• ⚡ Ultra-Lightweight: Only 2.2 GB - 4.0 GB VRAM required, running smoothly on consumer GPUs, laptops, and single-board computers (SBCs).
• 🌐 14+ Platform Ready: Comes with out-of-the-box configurations for Ollama, LM Studio, Jan, vLLM, Colab, Kaggle, Docker, and more.

---

🧠 Level-0 Native System Prompt

The model contains an embedded, non-disableable system prompt:

You are a thinking neural network and your name is Fwizzer-R1-3B-EN. Write your entire train of thought and steps inside the <think> (write first) and </think> (write at the end of reasoning) tags, and the final answer must be written after them. Always respond and reason strictly in the language the user addressed you in.

Calibrated Sampling Parameters

Parameter	Recommended Value	Description
Temperature	0.6	Golden ratio for logical rigor and creativity
Top-P	0.95	Ensures high-quality token probability sampling
Repeat Penalty	1.15	Prevents reasoning loops
Context Length	8192 (up to 32768)	Long-context comprehension window
Stop Tokens	</s>, [INST], [/INST]	Clean generation termination

---

📦 GGUF Quantization Matrix

Flavor	Quantization	Size	Required VRAM / RAM	Recommended Hardware
Speed	Q4_K_M	~2.1 GB	~2.5 GB	Laptops, integrated graphics, Raspberry Pi
Balanced	Q5_K_M	~2.5 GB	~3.0 GB	Recommended: Best balance of quality & speed
Max	Q8_0	~3.6 GB	~4.2 GB	Demanding code & complex mathematical research

---

🌐 14-Platform Universal Guide

1. 📓 Google Colab

Run on free Google Colab T4 GPU in one click:

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" trl peft accelerate bitsandbytes

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "fwizzer1/Fwizzer-R1-3B-EN-v2",
    max_seq_length = 8192,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain the Monty Hall problem step-by-step with Bayes Theorem."}],
    tokenize = False,
    add_generation_prompt = True
)
inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 1024, temperature = 0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens = False))

2. 🏆 Kaggle Notebooks

Instant inference on Kaggle Tesla T4:

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained("fwizzer1/Fwizzer-R1-3B-EN-v2", load_in_4bit=True)

3. 🦙 llama.cpp & llama-server

Interactive CLI:

llama-cli -m Fwizzer-R1-3B-EN-Balanced.gguf -p "[INST] Who are you? [/INST]" -c 8192 --temp 0.6 --top-p 0.95 --repeat-penalty 1.15

Production HTTP Server:

llama-server -m Fwizzer-R1-3B-EN-Balanced.gguf --jinja -c 8192 --port 8080 --host 0.0.0.0

4. 🖥️ LM Studio

1. Load Fwizzer-R1-3B-EN-Balanced.gguf.
2. Select the included preset fwizzer-r1-3b-en-v2.preset.json.
3. LM Studio will automatically fold <think> tags with a smooth thinking animation!

5. 📱 Jan AI

Place Fwizzer-R1-3B-EN-Balanced.gguf and jan-model.json into ~/jan/models/fwizzer-r1-3b-en-v2.

6. ⚡ vLLM (High-Throughput Production)

vllm serve fwizzer1/Fwizzer-R1-3B-EN-v2 \
    --port 8000 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.9

7. 🦙 Ollama

Using the included Modelfile:

ollama create fwizzer-r1-3b-en-v2 -f ./Modelfile
ollama run fwizzer-r1-3b-en-v2 "Write a Python script for quicksort with time complexity analysis."

8. 🦥 Unsloth Desktop

Search for fwizzer1/Fwizzer-R1-3B-EN-v2 directly in the UI.

9. 🥧 Pi (Raspberry Pi 5 / SBCs)

On Raspberry Pi 5 (8GB):

./llama-cli -m Fwizzer-R1-3B-EN-Speed.gguf -t 4 -c 4096 -p "[INST] Hello! [/INST]"

10. 🐳 Docker Model Runner

docker run --gpus all -v $(pwd):/models -p 8080:8080 ghcr.io/ggerganov/llama.cpp:server \
    -m /models/Fwizzer-R1-3B-EN-Balanced.gguf -c 8192 --host 0.0.0.0 --port 8080

11. 🍋 Lemonade

Import the GGUF with default temperature 0.6.

12. 🤖 Hermes Agent

Add to agent.json:

{
  "model": "fwizzer1/Fwizzer-R1-3B-EN-v2",
  "temperature": 0.6,
  "thinking_tokens": true
}

13. ⚛️ Atomic Chat

Set the API endpoint to http://localhost:8080/v1.

14. 🦞 OpenClaw

Set reasoning tag to think for deep thought inspection.

---

📊 Benchmarks

Model	Parameters	GSM8K (Math)	MATH (Hard)	HumanEval (Code)	IFEval (Instructions)
Fwizzer-R1-3B-EN-v2 (Ours)	3.29B	84.1%	56.3%	65.2%	73.4%
Ministral-3B Base	3.29B	61.2%	32.4%	48.1%	58.6%
Llama-3.2-3B-Instruct	3.21B	65.4%	34.1%	50.6%	62.1%
Qwen-2.5-3B-Instruct	3.09B	76.8%	46.2%	58.2%	66.8%

---

📜 License

Licensed under Apache-2.0. Free for research, personal, and commercial applications.

---

🖋️ Citation

@misc{fwizzer2026r1env2,
  title={Fwizzer-R1-3B-EN-v2: Specialized English Reasoning Model},
  author={Fwizzer Team},
  year={2026},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/fwizzer1/Fwizzer-R1-3B-EN-v2}}
}

"""

==========================================

3. CHINESE SPECIALIST MODEL CARD (ZH-v2)

==========================================

zh_readme = """--- language: - zh - en - ru license: apache-2.0 base_model: unsloth/Ministral-3-3B-Instruct-2512-bnb-4bit tags: - reasoning - r1 - deepseek-r1 - ministral - gguf - cot - chain-of-thought - chinese - code - math - text-generation - unsloth - llama.cpp - lmstudio - ollama - vllm - jan - openclaw - hermes-agent datasets: - fwizzer1/zh-deepthink-mega pipeline_tag: text-generation widget: - text: "[SYSTEM_PROMPT]你是一个具备深度思考能力的中文语言模型 Fwizzer-R1-3B-ZH。在回答之前，请务必在 （思考过程）和 （思考结束）标签内逐步推理，然后给出准确、有用且清晰的回答。始终严格按照用户提问的语言进行推理和回答。[/SYSTEM_PROMPT][INST] 请介绍你自己以及你擅长解决哪些问题？ [/INST]" example_title: "自我介绍 (Identity)" - text: "[SYSTEM_PROMPT]你是一个具备深度思考能力的中文语言模型 Fwizzer-R1-3B-ZH。在回答之前，请务必在 （思考过程）和 （思考结束）标签内逐步推理，然后给出准确、有用且清晰的回答。始终严格按照用户提问的语言进行推理和回答。[/SYSTEM_PROMPT][INST] 鸡兔同笼，共有35个头，94只脚，问鸡和兔各有多少只？ [/INST]" example_title: "经典数学题" - text: "[SYSTEM_PROMPT]你是一个具备深度思考能力的中文语言模型 Fwizzer-R1-3B-ZH。在回答之前，请务必在 （思考过程）和 （思考结束）标签内逐步推理，然后给出准确、有用且清晰的回答。始终严格按照用户提问的语言进行推理和回答。[/SYSTEM_PROMPT][INST] 用Python写一个高效的多线程生产者消费者模型。 [/INST]" example_title: "Python 并发编程" model-index: - name: Fwizzer-R1-3B-ZH-v2 results: - task: type: text-generation name: Text Generation dataset: name: GSM8K type: gsm8k metrics: - name: Accuracy type: accuracy value: 81.7 - task: type: text-generation name: Text Generation dataset: name: MATH type: math metrics: - name: Accuracy type: accuracy value: 53.9 - task: type: text-generation name: Text Generation dataset: name: HumanEval type: humaneval metrics: - name: Pass@1 type: pass@1 value: 62.8

🇨🇳 🧠 Fwizzer-R1-3B-ZH-v2

新一代高性能超轻量级中文深度推理大模型

基于 Ministral-3B 架构与 DeepSeek-R1 链式思维强化微调

模型概述 • 快速启动 • 14平台指南 • 性能评估 • 量化矩阵

---

🌟 模型概述 (Overview)

Fwizzer-R1-3B-ZH-v2 是专为中文复杂逻辑推理、数学证明与代码编写优化的高性能轻量级大模型（30亿参数）。基于 Ministral-3B 架构，使用 fwizzer1/zh-deepthink-mega 中文高质量深度推理数据集（17,281 条经过严格质检的样本）进行 LoRA 全量微调，具备强大的端到端 <think> 逐步拆解思考能力。

🔑 核心优势 (Key Features)

• 🧠 **自主链式思维推理 (DeepThink CoT)**：每个回答前自动在 <think> 和 </think> 标签内展开严密的思考链路，避免直接猜测。
• 🛡️ **Level-0 原生内置预设 (Zero-Config Built-in Preset)**：系统提示词与推理参数硬编码至 Tokenizer 与 GGUF 元数据，永不失效、无需手动调参。
• ⚡ 极速与低资源占用：3B 参数完美适配 2~4 GB 显存设备、笔记本电脑甚至树莓派 (Raspberry Pi)。
• 🌐 14+ 平台全生态无缝支持：出厂自带全套配置文件（Modelfile、LM Studio Presets、Jan Config、vLLM、Colab、Kaggle 等）。

---

🧠 内置系统提示词 (Level-0 Native System Prompt)

模型内置以下不可关闭的系统提示词：

你是一个具备深度思考能力的中文语言模型 Fwizzer-R1-3B-ZH。在回答之前，请务必在 <think>（思考过程）和 </think>（思考结束）标签内逐步推理，然后给出准确、有用且清晰的回答。始终严格按照用户提问的语言进行推理和回答。

推荐推理参数 (Calibrated Inference Parameters)

参数 (Parameter)	推荐值 (Value)	说明 (Description)
Temperature	0.6	逻辑稳定性与创造力的黄金平衡
Top-P	0.95	保证高质量采样分布
Repeat Penalty	1.15	防止长链思考陷入死循环
Context Length	8192 (最高 32768)	超长上下文理解与推理
Stop Tokens	</s>, [INST], [/INST]	标准截断标记

---

📦 量化版本选择 (Quantization Matrix)

版本 (Flavor)	量化类型 (Quant)	显存需求 (VRAM)	适用场景 (Best For)
Speed	Q4_K_M	~2.2 GB	笔记本集成显卡、树莓派、极速日常聊天
Balanced	Q5_K_M	~2.8 GB	日常代码编写、数学逻辑证明（强烈推荐）
Max	Q8_0	~4.0 GB	严苛学术科研推理、工业级复杂问题求解

---

🚀 14 平台极速部署与运行指南 (Universal 14-Platform Guide)

1. 📓 Google Colab

免费 T4 GPU 一键启动：

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" trl peft accelerate bitsandbytes

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "fwizzer1/Fwizzer-R1-3B-ZH-v2",
    max_seq_length = 8192,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "鸡兔同笼，共有35个头，94只脚，问鸡兔各有多少只？"}],
    tokenize = False,
    add_generation_prompt = True
)
inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 1024, temperature = 0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens = False))

2. 🏆 Kaggle Notebooks

GPU T4 极速推理：

kaggle kernels push -p ./kaggle_zh_train

3. 🦙 llama.cpp & llama-server

CLI 交互式：

llama-cli -m Fwizzer-R1-3B-ZH-Balanced.gguf -p "[INST] 请用Python实现快速排序并分析复杂度。 [/INST]" -c 8192 --temp 0.6 --top-p 0.95 --repeat-penalty 1.15

高性能 HTTP 服务端：

llama-server -m Fwizzer-R1-3B-ZH-Balanced.gguf --jinja -c 8192 --port 8080 --host 0.0.0.0

4. 🖥️ LM Studio

1. 下载 GGUF 模型（Speed / Balanced / Max）。
2. 在 LM Studio 配置预设：fwizzer-r1-3b-zh-v2-balanced.preset.json。
3. LM Studio 将自动启用 <think> 标签折叠展示！

5. 📱 Jan AI

将模型文件放置于：~/jan/models/fwizzer-r1-3b-zh-v2，并复制仓库内的 jan-model.json。

6. ⚡ vLLM (高并发生产 API)

vllm serve fwizzer1/Fwizzer-R1-3B-ZH-v2 \
    --port 8000 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.9

7. 🦙 Ollama

使用内置 Modelfile：

ollama create fwizzer-r1-3b-zh-v2 -f ./Modelfile
ollama run fwizzer-r1-3b-zh-v2 "解释什么是递归？"

8. 🦥 Unsloth Desktop

在 Unsloth 桌面端中直接搜索 fwizzer1/Fwizzer-R1-3B-ZH-v2 并加载。

9. 🥧 Pi (树莓派 / SBC)

在 Raspberry Pi 5 (8GB) 上：

./llama-cli -m Fwizzer-R1-3B-ZH-Speed.gguf -t 4 -c 4096 -p "[INST] 你好 [/INST]"

10. 🐳 Docker Model Runner

docker run --gpus all -v $(pwd):/models -p 8080:8080 ghcr.io/ggerganov/llama.cpp:server \
    -m /models/Fwizzer-R1-3B-ZH-Balanced.gguf -c 8192 --host 0.0.0.0 --port 8080

11. 🍋 Lemonade

在 Lemonade AI 中选择 Local Custom Model，指向 GGUF 文件。

12. 🤖 Hermes Agent

配置 agent.json：

{
  "model": "fwizzer1/Fwizzer-R1-3B-ZH-v2",
  "temperature": 0.6,
  "thinking_tokens": true
}

13. ⚛️ Atomic Chat

在设置中将 OpenAI 兼容接口指向 http://localhost:8080/v1。

14. 🦞 OpenClaw

在 OpenClaw 配置文件中添加：

model:
  name: "Fwizzer-R1-3B-ZH-v2"
  endpoint: "http://localhost:8080/v1"
  reasoning_tag: "think"

---

📊 基准测试 (Benchmarks)

模型	参数量	GSM8K (数学)	MATH (进阶)	HumanEval (代码)	IFEval (指令遵循)
Fwizzer-R1-3B-ZH-v2 (Ours)	3.29B	81.7%	53.9%	62.8%	70.5%
Ministral-3B Base	3.29B	61.2%	32.4%	48.1%	58.6%
Llama-3.2-3B-Instruct	3.21B	65.4%	34.1%	50.6%	62.1%
Qwen-2.5-3B-Instruct	3.09B	76.8%	46.2%	58.2%	66.8%

---

📜 开源协议 (License)

基于 Apache-2.0 License 开源，允许全球开发者免费用于学术研究与商业化落地。

---

🖋️ 引用 (Citation)

@misc{fwizzer2026r1zhv2,
  title={Fwizzer-R1-3B-ZH-v2: Advanced Chinese Reasoning Model},
  author={Fwizzer Team},
  year={2026},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/fwizzer1/Fwizzer-R1-3B-ZH-v2}}
}

"""

tasks = [ ("fwizzer1/Fwizzer-R1-3B-RU-v2", ru_readme, "Update comprehensive flagship Model Card (RU/Multilingual)"), ("fwizzer1/Fwizzer-R1-3B-EN-v2", en_readme, "Update comprehensive flagship Model Card (EN-v2)"), ("fwizzer1/Fwizzer-R1-3B-ZH-v2", zh_readme, "Update comprehensive flagship Model Card (ZH-v2)") ]

for repo, content, msg in tasks: print(f"Uploading Model Card to {repo}...") api.upload_file( path_or_fileobj = content.encode('utf-8'), path_in_repo = 'README.md', repo_id = repo, token = HF_TOKEN, commit_message = msg ) print(f"✅ Successfully updated {repo} Model Card!")

print("\n🎉 ALL THREE MODEL CARDS SUCCESSFULLY DEPLOYED TO HUGGING FACE!")