Hugging Face's logo

Mungert
/
Llama-3.2-3B-F1-Reasoning-Instruct-GGUF

Transformers
GGUF
English
Chinese
Taiwan
R.O.C
zhtw
SLM
Llama-32
Eval Results
imatrix
conversational
Model card Files
xet
Community

Llama-3.2-3B-F1-Reasoning-Instruct GGUF Models

Model Generation Details

This model was generated using llama.cpp at commit 064cc596.

Ultra-Low-Bit Quantization with IQ-DynamicGate (1-2 bit)

Our latest quantization method introduces precision-adaptive quantization for ultra-low-bit models (1-2 bit), with benchmark-proven improvements on Llama-3-8B. This approach uses layer-specific strategies to preserve accuracy while maintaining extreme memory efficiency.

Benchmark Context

All tests conducted on Llama-3-8B-Instruct using:

  • Standard perplexity evaluation pipeline
  • 2048-token context window
  • Same prompt set across all quantizations

Method

  • Dynamic Precision Allocation:
    • First/Last 25% of layers → IQ4_XS (selected layers)
    • Middle 50% → IQ2_XXS/IQ3_S (increase efficiency)
  • Critical Component Protection:
    • Embeddings/output layers use Q5_K
    • Reduces error propagation by 38% vs standard 1-2bit

Quantization Performance Comparison (Llama-3-8B)

Quantization Standard PPL DynamicGate PPL Δ PPL Std Size DG Size Δ Size Std Speed DG Speed
IQ2_XXS 11.30 9.84 -12.9% 2.5G 2.6G +0.1G 234s 246s
IQ2_XS 11.72 11.63 -0.8% 2.7G 2.8G +0.1G 242s 246s
IQ2_S 14.31 9.02 -36.9% 2.7G 2.9G +0.2G 238s 244s
IQ1_M 27.46 15.41 -43.9% 2.2G 2.5G +0.3G 206s 212s
IQ1_S 53.07 32.00 -39.7% 2.1G 2.4G +0.3G 184s 209s

Key:

  • PPL = Perplexity (lower is better)
  • Δ PPL = Percentage change from standard to DynamicGate
  • Speed = Inference time (CPU avx2, 2048 token context)
  • Size differences reflect mixed quantization overhead

Key Improvements:

  • 🔥 IQ1_M shows massive 43.9% perplexity reduction (27.46 → 15.41)
  • 🚀 IQ2_S cuts perplexity by 36.9% while adding only 0.2GB
  • ⚡ IQ1_S maintains 39.7% better accuracy despite 1-bit quantization

Tradeoffs:

  • All variants have modest size increases (0.1-0.3GB)
  • Inference speeds remain comparable (<5% difference)

When to Use These Models

📌 Fitting models into GPU VRAM

✔ Memory-constrained deployments

✔ Cpu and Edge Devices where 1-2bit errors can be tolerated

✔ Research into ultra-low-bit quantization

Choosing the Right Model Format

Selecting the correct model format depends on your hardware capabilities and memory constraints.

BF16 (Brain Float 16) – Use if BF16 acceleration is available

  • A 16-bit floating-point format designed for faster computation while retaining good precision.
  • Provides similar dynamic range as FP32 but with lower memory usage.
  • Recommended if your hardware supports BF16 acceleration (check your device's specs).
  • Ideal for high-performance inference with reduced memory footprint compared to FP32.

📌 Use BF16 if:
✔ Your hardware has native BF16 support (e.g., newer GPUs, TPUs).
✔ You want higher precision while saving memory.
✔ You plan to requantize the model into another format.

📌 Avoid BF16 if:
❌ Your hardware does not support BF16 (it may fall back to FP32 and run slower).
❌ You need compatibility with older devices that lack BF16 optimization.

F16 (Float 16) – More widely supported than BF16

  • A 16-bit floating-point high precision but with less of range of values than BF16.
  • Works on most devices with FP16 acceleration support (including many GPUs and some CPUs).
  • Slightly lower numerical precision than BF16 but generally sufficient for inference.

📌 Use F16 if:
✔ Your hardware supports FP16 but not BF16.
✔ You need a balance between speed, memory usage, and accuracy.
✔ You are running on a GPU or another device optimized for FP16 computations.

📌 Avoid F16 if:
❌ Your device lacks native FP16 support (it may run slower than expected).
❌ You have memory limitations.

Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference

Quantization reduces model size and memory usage while maintaining as much accuracy as possible.

  • Lower-bit models (Q4_K) → Best for minimal memory usage, may have lower precision.
  • Higher-bit models (Q6_K, Q8_0) → Better accuracy, requires more memory.

📌 Use Quantized Models if:
✔ You are running inference on a CPU and need an optimized model.
✔ Your device has low VRAM and cannot load full-precision models.
✔ You want to reduce memory footprint while keeping reasonable accuracy.

📌 Avoid Quantized Models if:
❌ You need maximum accuracy (full-precision models are better for this).
❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).

Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)

These models are optimized for extreme memory efficiency, making them ideal for low-power devices or large-scale deployments where memory is a critical constraint.

  • IQ3_XS: Ultra-low-bit quantization (3-bit) with extreme memory efficiency.

    • Use case: Best for ultra-low-memory devices where even Q4_K is too large.
    • Trade-off: Lower accuracy compared to higher-bit quantizations.

  • IQ3_S: Small block size for maximum memory efficiency.

    • Use case: Best for low-memory devices where IQ3_XS is too aggressive.

  • IQ3_M: Medium block size for better accuracy than IQ3_S.

    • Use case: Suitable for low-memory devices where IQ3_S is too limiting.

  • Q4_K: 4-bit quantization with block-wise optimization for better accuracy.

    • Use case: Best for low-memory devices where Q6_K is too large.

  • Q4_0: Pure 4-bit quantization, optimized for ARM devices.

    • Use case: Best for ARM-based devices or low-memory environments.

Summary Table: Model Format Selection

Model Format Precision Memory Usage Device Requirements Best Use Case
BF16 Highest High BF16-supported GPU/CPUs High-speed inference with reduced memory
F16 High High FP16-supported devices GPU inference when BF16 isn't available
Q4_K Medium Low Low CPU or Low-VRAM devices Best for memory-constrained environments
Q6_K Medium Moderate CPU with more memory Better accuracy while still being quantized
Q8_0 High Moderate CPU or GPU with enough VRAM Best accuracy among quantized models
IQ3_XS Very Low Very Low Ultra-low-memory devices Extreme memory efficiency and low accuracy
Q4_0 Low Low ARM or low-memory devices llama.cpp can optimize for ARM devices

Included Files & Details

Llama-3.2-3B-F1-Reasoning-Instruct-bf16.gguf

  • Model weights preserved in BF16.
  • Use this if you want to requantize the model into a different format.
  • Best if your device supports BF16 acceleration.

Llama-3.2-3B-F1-Reasoning-Instruct-f16.gguf

  • Model weights stored in F16.
  • Use if your device supports FP16, especially if BF16 is not available.

Llama-3.2-3B-F1-Reasoning-Instruct-bf16-q8_0.gguf

  • Output & embeddings remain in BF16.
  • All other layers quantized to Q8_0.
  • Use if your device supports BF16 and you want a quantized version.

Llama-3.2-3B-F1-Reasoning-Instruct-f16-q8_0.gguf

  • Output & embeddings remain in F16.
  • All other layers quantized to Q8_0.

Llama-3.2-3B-F1-Reasoning-Instruct-q4_k.gguf

  • Output & embeddings quantized to Q8_0.
  • All other layers quantized to Q4_K.
  • Good for CPU inference with limited memory.

Llama-3.2-3B-F1-Reasoning-Instruct-q4_k_s.gguf

  • Smallest Q4_K variant, using less memory at the cost of accuracy.
  • Best for very low-memory setups.

Llama-3.2-3B-F1-Reasoning-Instruct-q6_k.gguf

  • Output & embeddings quantized to Q8_0.
  • All other layers quantized to Q6_K .

Llama-3.2-3B-F1-Reasoning-Instruct-q8_0.gguf

  • Fully Q8 quantized model for better accuracy.
  • Requires more memory but offers higher precision.

Llama-3.2-3B-F1-Reasoning-Instruct-iq3_xs.gguf

  • IQ3_XS quantization, optimized for extreme memory efficiency.
  • Best for ultra-low-memory devices.

Llama-3.2-3B-F1-Reasoning-Instruct-iq3_m.gguf

  • IQ3_M quantization, offering a medium block size for better accuracy.
  • Suitable for low-memory devices.

Llama-3.2-3B-F1-Reasoning-Instruct-q4_0.gguf

  • Pure Q4_0 quantization, optimized for ARM devices.
  • Best for low-memory environments.
  • Prefer IQ4_NL for better accuracy.

🚀 If you find these models useful

❤ Please click "Like" if you find this useful!
Help me test my AI-Powered Network Monitor Assistant with quantum-ready security checks:
👉 Free Network Monitor

💬 How to test:
Choose an AI assistant type:

  • TurboLLM (GPT-4o-mini)
  • HugLLM (Hugginface Open-source)
  • TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

  • Function calling against live network services
  • How small can a model go while still handling:
    • Automated Nmap scans
    • Quantum-readiness checks
    • Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads):

  • ✅ Zero-configuration setup
  • ⏳ 30s load time (slow inference but no API costs)
  • 🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4o-mini for:

🔵 HugLLM – Latest Open-source models:

  • 🌐 Runs on Hugging Face Inference API

💡 Example commands to you could test:

  1. "Give me info on my websites SSL certificate"
  2. "Check if my server is using quantum safe encyption for communication"
  3. "Run a comprehensive security audit on my server"
  4. '"Create a cmd processor to .. (what ever you want)" Note you need to install a Free Network Monitor Agent to run the .net code from. This is a very flexible and powerful feature. Use with caution!

Model Card for Llama-3.2-3B-F1-Reasoning-Instruct (a.k.a Formosa-1-Reasoning or F1-Reasoning)

image/png

Llama-3.2-3B-F1-Reasoning-Instruct（a.k.a Formosa-1-Reasoning or F1-Reasoning） 是由 Twinkle AI 與 APMIC 合作開發，並在國家高速網路與計算中心技術指導之下，針對中華民國台灣語境與任務需求所微調之繁體中文語言模型，涵蓋法律、教育、生活應用等多元場景，並以高指令跟隨能力為目標進行強化。

Model Details

Model Description

Model Sources

Evaluation

Results

下表採用 🌟 Twinkle Eval 評測框架

模型 評測模式 TMMLU+(%) 台灣法律(%) MMLU(%) 測試次數 選項排序
mistralai/Mistral-Small-24B-Instruct-2501 box 56.15 (±0.0172) 37.48 (±0.0098) 74.61 (±0.0154) 3 隨機
meta-llama/Llama-3.2-3B-Instruct box 15.49 (±0.0104) 25.68 (±0.0200) 6.90 (±0.0096) 3 隨機
meta-llama/Llama-3.2-3B-Instruct pattern 35.85 (±0.0174) 32.22 (±0.0023) 59.33 (±0.0168) 3 隨機
MediaTek-Research/Llama-Breeze2-3B-Instruct pattern 40.32 (±0.0181) 38.92 (±0.0193) 55.37 (±0.0180) 3 隨機
twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct (ours) box 46.16 (±0.0198) 34.92 (±0.0243) 51.22 (±0.0206) 3 隨機

下表用 lighteval 評測框架

模型 MATH-500 GPQA Diamond
meta-llama/Llama-3.2-3B-Instruct 44.40 27.78
twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct (ours) 51.40 33.84

🔧 Tool Calling

本模型使用 Hermes 格式訓練，並支援平行呼叫（Parallel calling），以下為完整範例流程。 Tool call 模板已經為大家寫好放進 chat-template 了，Enjoy it！

1️⃣ 啟動 vLLM 後端

⚠️ 注意：需要 vLLM 版本 >= 0.8.3，否則 enable-reasoning、enable-auto-tool-choice 無法同時開啟

vllm serve twinkle-ai/Llama-3.2-3B-F1-Reasoning-Instruct \
  --port 8001 \
  --enable-reasoning \
  --reasoning-parser deepseek_r1 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

2️⃣ 定義工具（Functions） 

def get_weather(location: str, unit: str):
    return f"{location}的氣溫是{unit}26度，晴朗無風"

def search(query: str):
    return "川普終於宣布對等關稅政策，針對 18 個經濟體課徵一半的對等關稅，並從 4/5 起對所有進口產品徵收10%的基準關稅！美國將針對被認定為不當貿易行為(不公平貿易) 的國家，於 4/9 起課徵報復型對等關稅 (Discounted Reciprocal Tariff)，例如：日本將被課徵 24% 的關稅，歐盟則為 20%，以取代普遍性的 10% 關稅。\n針對中國則開啟新一波 34% 關稅，並疊加於先前已實施的關稅上，這將使中國進口商品的基本關稅稅率達到 54%，而且這尚未包含拜登總統任內或川普第一任期所施加的額外關稅。加拿大與墨西哥則不適用這套對等關稅制度，但川普認為這些國家在芬太尼危機與非法移民問題尚未完全解決，因此計畫對這兩國的大多數進口商品施加 25% 關稅。另外原本針對汽車與多數其他商品的關稅豁免將於 4/2 到期。\n台灣的部分，美國擬向台灣課徵32％的對等關稅，雖然並未針對晶片特別課徵關稅，但仍在記者會中提到台灣搶奪所有的電腦與半導體晶片，最終促成台積電對美國投資計劃額外加碼 1,000 億美元的歷史性投資；歐盟則課徵20％的對等關稅。最後是汽車關稅將於 4/2 起，對所有外國製造的汽車課徵25% 關稅。"

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather in a given location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "國家或城市名, e.g., 'Taipei'、'Jaipei'"},
                    "unit": {"type": "string", "description": "氣溫單位，亞洲城市使用攝氏；歐美城市使用華氏", "enum": ["celsius", "fahrenheit"]}
                },
                "required": ["location", "unit"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "search",
            "description": "這是一個類似 Google 的搜尋引擎，關於知識、天氣、股票、電影、小說、百科等等問題，如果你不確定答案就搜尋一下。",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "should be a search query, e.g., '2024 南韓 戒嚴'"}
                },
                "required": ["query"]
            }
        }
    }
]

3️⃣ 執行工具調用（Tool Calls）

⚠️ 注意：system_prompt 可以不用帶，除非是需要時間基準的工具。

response = client.chat.completions.create(
    model=client.models.list().data[0].id,
    messages=[
        {"role": "system", "content": "記住你的知識截止於 2024/12，今天是 2025/4/7"},
        {"role": "user", "content": "台北氣溫如何? 另外，告訴我川普最新關稅政策"},
    ],
    max_tokens=1500,
    temperature=0.6,
    top_p=0.95,
    tools=tools,
    tool_choice="auto"
)

print(response.choices[0].message.reasoning_content)
print(response.choices[0].message.tool_calls)

🧠 推理內容輸出（僅顯示部分）

好的，我需要幫助這個使用者解決他們的問題。他們問了兩件事：首先，臺北市的天氣情況，以及第二，關於川普最近的關稅政策。
對於第一部分，他們提到了“臺北”，所以應該呼叫 get_weather 函式…
接下來是關於川普的新關稅政策…
總結一下，我需要分別進行兩次 API 呼叫，每次都有各自正確填寫的參數…

⚙️ Tool Calls List 

[ChatCompletionMessageToolCall(id='chatcmpl-tool-35e74420119349999913a10133b84bd3', function=Function(arguments='{"location": "Taipei", "unit": "celsius"}', name='get_weather'), type='function'), ChatCompletionMessageToolCall(id='chatcmpl-tool-7ffdcb98e59f4134a6171defe7f2e31b', function=Function(arguments='{"query": "Donald Trump latest tariffs policy"}', name='search'), type='function')]

4️⃣ 產生最終回答 

response = client.chat.completions.create(
    model=client.models.list().data[0].id,
    messages=[
        {"role": "system", "content": "記住你的知識截止於 2024/12，今天是 2025/4/7"},
        {"role": "user", "content": "台北氣溫如何? 另外，告訴我川普最新關稅政策"},
        {
            "role": "assistant",
            "content": "",
            "tool_calls": [
                {
                    "id": response.choices[0].message.tool_calls[0].id,
                    "type": "function",
                    "function": {
                        "name": response.choices[0].message.tool_calls[0].function.name,
                        "arguments": response.choices[0].message.tool_calls[0].function.arguments
                    }
                },
                {
                    "id": response.choices[0].message.tool_calls[1].id,
                    "type": "function",
                    "function": {
                        "name": response.choices[0].message.tool_calls[1].function.name,
                        "arguments": response.choices[0].message.tool_calls[1].function.arguments
                    }
                }
            ]
        },
        {
            "role": "tool",
            "content": search(**json.loads(response.choices[0].message.tool_calls[0].function.arguments)),
            "tool_call_id": response.choices[0].message.tool_calls[0].id # tool_call_id 必須要帶，才能正確配對 工具 及 tool_call
        },
        {
            "role": "tool",
            "content": get_weather(**json.loads(response.choices[0].message.tool_calls[1].function.arguments)),
            "tool_call_id": response.choices[0].message.tool_calls[1].id # tool_call_id 必須要帶，才能正確配對 工具 及 tool_call
        }
    ],
    max_tokens=1500,
    temperature=0.6,
    top_p=0.95,
    tools=tools,
    tool_choice="auto"
)

print(response.choices[0].message.reasoning_content)
print(response.choices[0].message.content)

🧠 推理內容輸出（僅顯示部分）

首先，我需要處理使用者的查詢，他們要求了解臺北市的當下氣溫以及川普最近的關稅政策…
在呼叫了 get_weather 後得到了臺北市氣溫為 26 度（攝氏）…
接著，使用 search 搜尋「川普最新關稅政策 2025」…
整合後，我提供如下摘要：

📋 最終輸出內容 

以下是您請求的資訊：

**臺北市氣溫**
- 目前的氣溫為 **26°C**（攝氏）
- 天候狀況：晴朗無風

**川普最新關稅政策概述**
1. **對等關稅政策**  
   - 對 18 個經濟體課徵 50% 的對等關稅  
   - 自 4 月 5 日起，所有進口產品全面徵收 10% 基本關稅  

2. **報復型對等關稅**  
   - 日本 24%、歐盟 20%  

3. **對中國的高額關稅**  
   - 增加至 54%（原有關稅 + 新增 34%）  

4. **特殊案例**  
   - 加拿大與墨西哥不適用，但其他商品課徵 25%  
   - 汽車與部分商品的免稅即將到期  

5. **對台灣的影響**  
   - 美國計畫對台灣課徵 32% 關稅，但晶片暫無額外課稅  

6. **全球視角**  
   - 歐盟與日本關稅比例相對較高

Citation 

@misc{twinkleai2025llama3.2f1,
  title        = {Llama-3.2-3B-F1-Reasoning-Instruct: A Traditional Chinese Instruction-Tuned Reasoning Language Model for Taiwan},
  author       = {Huang, Liang Hsun and Chen, Min Yi and Lin, Wen Bin and Chuang, Chao Chun and Sung, Dave},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/twinkle-ai/Llama-3.2-3B-F1-Instruct}},
  note         = {Twinkle AI and APMIC. All authors contributed equally.}
}

Acknowledge

  • 特此感謝國家高速網路與計算中心的指導與 APMIC 的算力支援，才得以讓本專案訓利完成。
  • 特此致謝黃啟聖老師、許武龍（哈爸）、臺北市立第一女子高級中學物理科陳姿燁老師、奈視科技 CTO Howard、AIPLUX Technology、郭家嘉老師以及所有在資料集製作過程中提供寶貴協助的夥伴。

Model Card Authors

Twinkle AI

Model Card Contact

Twinkle AI

Downloads last month
166
GGUF
Model size
3.61B params
Architecture
llama
Hardware compatibility
Log In to view the estimation

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

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

Model tree for Mungert/Llama-3.2-3B-F1-Reasoning-Instruct-GGUF

Base model

meta-llama/Llama-3.2-3B
Finetuned
lianghsun/Llama-3.2-3B-F1-Base
Quantized
(3)
this model

Datasets used to train Mungert/Llama-3.2-3B-F1-Reasoning-Instruct-GGUF

Evaluation results

View on Papers With Code