mungert/llama-3_3-nemotron-super-49b-v1-gguf overview

{
  "metadata": {},
  "card_data": {
    "library_name": "transformers",
    "license": "other",
    "license_name": "nvidia-open-model-license",
    "license_link": "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/",
    "pipeline_tag": "text-generation",
    "language": [
      "en"
    ],
    "tags": [
      "nvidia",
      "llama-3",
      "pytorch"
    ],
    "frontmatter": {
      "library_name": "transformers",
      "license": "other",
      "license_name": "nvidia-open-model-license",
      "license_link": ">-",
      "pipeline_tag": "text-generation",
      "language": [
        "en"
      ],
      "tags": [
        "nvidia",
        "llama-3",
        "pytorch"
      ]
    },
    "hero_image_url": "flow.png",
    "summary": "",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlibrary_name: transformers\nlicense: other\nlicense_name: nvidia-open-model-license\nlicense_link: >-\n  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/\n\npipeline_tag: text-generation\nlanguage:\n  - en\ntags:\n  - nvidia\n  - llama-3\n  - pytorch\n---\n\n# <span style=\"color: #7FFF7F;\">Llama-3_3-Nemotron-Super-49B-v1 GGUF Models</span>\n\n## <span style=\"color: #7FFF7F;\">Ultra-Low-Bit Quantization with IQ-DynamicGate (1-2 bit)</span>\n\nOur 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.\n\n### **Benchmark Context**\nAll tests conducted on **Llama-3-8B-Instruct** using:\n- Standard perplexity evaluation pipeline\n- 2048-token context window\n- Same prompt set across all quantizations\n\n### **Method**\n- **Dynamic Precision Allocation**:  \n  - First/Last 25% of layers → IQ4_XS (selected layers)  \n  - Middle 50% → IQ2_XXS/IQ3_S (increase efficiency)  \n- **Critical Component Protection**:  \n  - Embeddings/output layers use Q5_K  \n  - Reduces error propagation by 38% vs standard 1-2bit  \n\n### **Quantization Performance Comparison (Llama-3-8B)**\n\n| Quantization | Standard PPL | DynamicGate PPL | Δ PPL   | Std Size | DG Size | Δ Size | Std Speed | DG Speed |\n|--------------|--------------|------------------|---------|----------|---------|--------|-----------|----------|\n| IQ2_XXS      | 11.30        | 9.84             | -12.9%  | 2.5G     | 2.6G    | +0.1G  | 234s      | 246s     |\n| IQ2_XS       | 11.72        | 11.63            | -0.8%   | 2.7G     | 2.8G    | +0.1G  | 242s      | 246s     |\n| IQ2_S        | 14.31        | 9.02             | -36.9%  | 2.7G     | 2.9G    | +0.2G  | 238s      | 244s     |\n| IQ1_M        | 27.46        | 15.41            | -43.9%  | 2.2G     | 2.5G    | +0.3G  | 206s      | 212s     |\n| IQ1_S        | 53.07        | 32.00            | -39.7%  | 2.1G     | 2.4G    | +0.3G  | 184s      | 209s     |\n\n**Key**:\n- PPL = Perplexity (lower is better)\n- Δ PPL = Percentage change from standard to DynamicGate\n- Speed = Inference time (CPU avx2, 2048 token context)\n- Size differences reflect mixed quantization overhead\n\n**Key Improvements:**\n- 🔥 **IQ1_M** shows massive 43.9% perplexity reduction (27.46 → 15.41)\n- 🚀 **IQ2_S** cuts perplexity by 36.9% while adding only 0.2GB\n- ⚡ **IQ1_S** maintains 39.7% better accuracy despite 1-bit quantization\n\n**Tradeoffs:**\n- All variants have modest size increases (0.1-0.3GB)\n- Inference speeds remain comparable (<5% difference)\n\n\n### **When to Use These Models**\n📌 **Fitting models into GPU VRAM**\n\n✔ **Memory-constrained deployments**\n\n✔ **Cpu and Edge Devices** where 1-2bit errors can be tolerated \n \n✔ **Research** into ultra-low-bit quantization\n\n\n## **Choosing the Right Model Format**  \n\nSelecting the correct model format depends on your **hardware capabilities** and **memory constraints**.  \n\n### **BF16 (Brain Float 16) – Use if BF16 acceleration is available**  \n- A 16-bit floating-point format designed for **faster computation** while retaining good precision.  \n- Provides **similar dynamic range** as FP32 but with **lower memory usage**.  \n- Recommended if your hardware supports **BF16 acceleration** (check your device's specs).  \n- Ideal for **high-performance inference** with **reduced memory footprint** compared to FP32.  \n\n📌 **Use BF16 if:**  \n✔ Your hardware has native **BF16 support** (e.g., newer GPUs, TPUs).  \n✔ You want **higher precision** while saving memory.  \n✔ You plan to **requantize** the model into another format.  \n\n📌 **Avoid BF16 if:**  \n❌ Your hardware does **not** support BF16 (it may fall back to FP32 and run slower).  \n❌ You need compatibility with older devices that lack BF16 optimization.  \n\n---\n\n### **F16 (Float 16) – More widely supported than BF16**  \n- A 16-bit floating-point **high precision** but with less of range of values than BF16. \n- Works on most devices with **FP16 acceleration support** (including many GPUs and some CPUs).  \n- Slightly lower numerical precision than BF16 but generally sufficient for inference.  \n\n📌 **Use F16 if:**  \n✔ Your hardware supports **FP16** but **not BF16**.  \n✔ You need a **balance between speed, memory usage, and accuracy**.  \n✔ You are running on a **GPU** or another device optimized for FP16 computations.  \n\n📌 **Avoid F16 if:**  \n❌ Your device lacks **native FP16 support** (it may run slower than expected).  \n❌ You have memory limitations.  \n\n---\n\n### **Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference**  \nQuantization reduces model size and memory usage while maintaining as much accuracy as possible.  \n- **Lower-bit models (Q4_K)** → **Best for minimal memory usage**, may have lower precision.  \n- **Higher-bit models (Q6_K, Q8_0)** → **Better accuracy**, requires more memory.  \n\n📌 **Use Quantized Models if:**  \n✔ You are running inference on a **CPU** and need an optimized model.  \n✔ Your device has **low VRAM** and cannot load full-precision models.  \n✔ You want to reduce **memory footprint** while keeping reasonable accuracy.  \n\n📌 **Avoid Quantized Models if:**  \n❌ You need **maximum accuracy** (full-precision models are better for this).  \n❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).  \n\n---\n\n### **Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)**  \nThese models are optimized for **extreme memory efficiency**, making them ideal for **low-power devices** or **large-scale deployments** where memory is a critical constraint.  \n\n- **IQ3_XS**: Ultra-low-bit quantization (3-bit) with **extreme memory efficiency**.  \n  - **Use case**: Best for **ultra-low-memory devices** where even Q4_K is too large.  \n  - **Trade-off**: Lower accuracy compared to higher-bit quantizations.  \n\n- **IQ3_S**: Small block size for **maximum memory efficiency**.  \n  - **Use case**: Best for **low-memory devices** where **IQ3_XS** is too aggressive.  \n\n- **IQ3_M**: Medium block size for better accuracy than **IQ3_S**.  \n  - **Use case**: Suitable for **low-memory devices** where **IQ3_S** is too limiting.  \n\n- **Q4_K**: 4-bit quantization with **block-wise optimization** for better accuracy.  \n  - **Use case**: Best for **low-memory devices** where **Q6_K** is too large.  \n\n- **Q4_0**: Pure 4-bit quantization, optimized for **ARM devices**.  \n  - **Use case**: Best for **ARM-based devices** or **low-memory environments**.  \n\n---\n\n### **Summary Table: Model Format Selection**  \n\n| Model Format  | Precision  | Memory Usage  | Device Requirements  | Best Use Case  |  \n|--------------|------------|---------------|----------------------|---------------|  \n| **BF16**     | Highest    | High          | BF16-supported GPU/CPUs  | High-speed inference with reduced memory |  \n| **F16**      | High       | High          | FP16-supported devices | GPU inference when BF16 isn't available |  \n| **Q4_K**     | Medium Low | Low           | CPU or Low-VRAM devices | Best for memory-constrained environments |  \n| **Q6_K**     | Medium     | Moderate      | CPU with more memory | Better accuracy while still being quantized |  \n| **Q8_0**     | High       | Moderate      | CPU or GPU with enough VRAM | Best accuracy among quantized models |  \n| **IQ3_XS**   | Very Low   | Very Low      | Ultra-low-memory devices | Extreme memory efficiency and low accuracy |  \n| **Q4_0**     | Low        | Low           | ARM or low-memory devices | llama.cpp can optimize for ARM devices |  \n\n---\n\n## **Included Files & Details**  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-bf16.gguf`  \n- Model weights preserved in **BF16**.  \n- Use this if you want to **requantize** the model into a different format.  \n- Best if your device supports **BF16 acceleration**.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-f16.gguf`  \n- Model weights stored in **F16**.  \n- Use if your device supports **FP16**, especially if BF16 is not available.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-bf16-q8_0.gguf`  \n- **Output & embeddings** remain in **BF16**.  \n- All other layers quantized to **Q8_0**.  \n- Use if your device supports **BF16** and you want a quantized version.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-f16-q8_0.gguf`  \n- **Output & embeddings** remain in **F16**.  \n- All other layers quantized to **Q8_0**.    \n\n### `Llama-3_3-Nemotron-Super-49B-v1-q4_k.gguf`  \n- **Output & embeddings** quantized to **Q8_0**.  \n- All other layers quantized to **Q4_K**.  \n- Good for **CPU inference** with limited memory.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-q4_k_s.gguf`  \n- Smallest **Q4_K** variant, using less memory at the cost of accuracy.  \n- Best for **very low-memory setups**.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-q6_k.gguf`  \n- **Output & embeddings** quantized to **Q8_0**.  \n- All other layers quantized to **Q6_K** .  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-q8_0.gguf`  \n- Fully **Q8** quantized model for better accuracy.  \n- Requires **more memory** but offers higher precision.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-iq3_xs.gguf`  \n- **IQ3_XS** quantization, optimized for **extreme memory efficiency**.  \n- Best for **ultra-low-memory devices**.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-iq3_m.gguf`  \n- **IQ3_M** quantization, offering a **medium block size** for better accuracy.  \n- Suitable for **low-memory devices**.  \n\n### `Llama-3_3-Nemotron-Super-49B-v1-q4_0.gguf`  \n- Pure **Q4_0** quantization, optimized for **ARM devices**.  \n- Best for **low-memory environments**.\n- Prefer IQ4_NL for better accuracy.\n\n# <span id=\"testllm\" style=\"color: #7F7FFF;\">🚀 If you find these models useful</span>\n\nPlease click like ❤ . Also I'd really appreciate it if you could test my Network Monitor Assistant at 👉 [Network Monitor Assitant](https://readyforquantum.com).\n\n💬 Click the **chat icon** (bottom right of the main and dashboard pages) . Choose a LLM; toggle between the LLM Types TurboLLM -> FreeLLM -> TestLLM.\n\n### What I'm Testing\n\nI'm experimenting with **function calling** against my network monitoring service. Using small open source models. I am into the question \"How small can it go and still function\".\n\n🟡 **TestLLM** – Runs the current testing model using llama.cpp on 6 threads of a Cpu VM (Should take about 15s to load. Inference speed is quite slow and it only processes one user prompt at a time—still working on scaling!). If you're curious, I'd be happy to share how it works! .\n\n### The other Available AI Assistants\n\n🟢 **TurboLLM** – Uses **gpt-4o-mini** Fast! . Note: tokens are limited since OpenAI models are pricey, but you can [Login](https://readyforquantum.com) or [Download](https://readyforquantum.com/download/?utm_source=huggingface&utm_medium=referral&utm_campaign=huggingface_repo_readme) the Quantum Network Monitor agent to get more tokens, Alternatively use the TestLLM .\n\n🔵 **HugLLM** – Runs **open-source Hugging Face models** Fast, Runs small models (≈8B) hence lower quality, Get 2x more tokens (subject to Hugging Face API availability)\n\n### Final Word\n\nI fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is [open source](https://github.com/Mungert69). Feel free to use whatever you find helpful.\n\nIf you appreciate the work, please consider [buying me a coffee](https://www.buymeacoffee.com/mahadeva) ☕. Your support helps cover service costs and allows me to raise token limits for everyone.\n\nI'm also open to job opportunities or sponsorship.\n\nThank you! 😊\n\n\n\n# <span style=\"color: #7FFF7F;\">Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF GGUF Models</span>\n\n## <span style=\"color: #7FFF7F;\">Ultra-Low-Bit Quantization with IQ-DynamicGate (1-2 bit)</span>\n\nOur 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.\n\n### **Benchmark Context**\nAll tests conducted on **Llama-3-8B-Instruct** using:\n- Standard perplexity evaluation pipeline\n- 2048-token context window\n- Same prompt set across all quantizations\n\n### **Method**\n- **Dynamic Precision Allocation**:  \n  - First/Last 25% of layers → IQ4_XS (selected layers)  \n  - Middle 50% → IQ2_XXS/IQ3_S (increase efficiency)  \n- **Critical Component Protection**:  \n  - Embeddings/output layers use Q5_K  \n  - Reduces error propagation by 38% vs standard 1-2bit  \n\n### **Quantization Performance Comparison (Llama-3-8B)**\n\n| Quantization | Standard PPL | DynamicGate PPL | Δ PPL   | Std Size | DG Size | Δ Size | Std Speed | DG Speed |\n|--------------|--------------|------------------|---------|----------|---------|--------|-----------|----------|\n| IQ2_XXS      | 11.30        | 9.84             | -12.9%  | 2.5G     | 2.6G    | +0.1G  | 234s      | 246s     |\n| IQ2_XS       | 11.72        | 11.63            | -0.8%   | 2.7G     | 2.8G    | +0.1G  | 242s      | 246s     |\n| IQ2_S        | 14.31        | 9.02             | -36.9%  | 2.7G     | 2.9G    | +0.2G  | 238s      | 244s     |\n| IQ1_M        | 27.46        | 15.41            | -43.9%  | 2.2G     | 2.5G    | +0.3G  | 206s      | 212s     |\n| IQ1_S        | 53.07        | 32.00            | -39.7%  | 2.1G     | 2.4G    | +0.3G  | 184s      | 209s     |\n\n**Key**:\n- PPL = Perplexity (lower is better)\n- Δ PPL = Percentage change from standard to DynamicGate\n- Speed = Inference time (CPU avx2, 2048 token context)\n- Size differences reflect mixed quantization overhead\n\n**Key Improvements:**\n- 🔥 **IQ1_M** shows massive 43.9% perplexity reduction (27.46 → 15.41)\n- 🚀 **IQ2_S** cuts perplexity by 36.9% while adding only 0.2GB\n- ⚡ **IQ1_S** maintains 39.7% better accuracy despite 1-bit quantization\n\n**Tradeoffs:**\n- All variants have modest size increases (0.1-0.3GB)\n- Inference speeds remain comparable (<5% difference)\n\n\n### **When to Use These Models**\n📌 **Fitting models into GPU VRAM**\n\n✔ **Memory-constrained deployments**\n\n✔ **Cpu and Edge Devices** where 1-2bit errors can be tolerated \n \n✔ **Research** into ultra-low-bit quantization\n\n\n## **Choosing the Right Model Format**  \n\nSelecting the correct model format depends on your **hardware capabilities** and **memory constraints**.  \n\n### **BF16 (Brain Float 16) – Use if BF16 acceleration is available**  \n- A 16-bit floating-point format designed for **faster computation** while retaining good precision.  \n- Provides **similar dynamic range** as FP32 but with **lower memory usage**.  \n- Recommended if your hardware supports **BF16 acceleration** (check your device's specs).  \n- Ideal for **high-performance inference** with **reduced memory footprint** compared to FP32.  \n\n📌 **Use BF16 if:**  \n✔ Your hardware has native **BF16 support** (e.g., newer GPUs, TPUs).  \n✔ You want **higher precision** while saving memory.  \n✔ You plan to **requantize** the model into another format.  \n\n📌 **Avoid BF16 if:**  \n❌ Your hardware does **not** support BF16 (it may fall back to FP32 and run slower).  \n❌ You need compatibility with older devices that lack BF16 optimization.  \n\n---\n\n### **F16 (Float 16) – More widely supported than BF16**  \n- A 16-bit floating-point **high precision** but with less of range of values than BF16. \n- Works on most devices with **FP16 acceleration support** (including many GPUs and some CPUs).  \n- Slightly lower numerical precision than BF16 but generally sufficient for inference.  \n\n📌 **Use F16 if:**  \n✔ Your hardware supports **FP16** but **not BF16**.  \n✔ You need a **balance between speed, memory usage, and accuracy**.  \n✔ You are running on a **GPU** or another device optimized for FP16 computations.  \n\n📌 **Avoid F16 if:**  \n❌ Your device lacks **native FP16 support** (it may run slower than expected).  \n❌ You have memory limitations.  \n\n---\n\n### **Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference**  \nQuantization reduces model size and memory usage while maintaining as much accuracy as possible.  \n- **Lower-bit models (Q4_K)** → **Best for minimal memory usage**, may have lower precision.  \n- **Higher-bit models (Q6_K, Q8_0)** → **Better accuracy**, requires more memory.  \n\n📌 **Use Quantized Models if:**  \n✔ You are running inference on a **CPU** and need an optimized model.  \n✔ Your device has **low VRAM** and cannot load full-precision models.  \n✔ You want to reduce **memory footprint** while keeping reasonable accuracy.  \n\n📌 **Avoid Quantized Models if:**  \n❌ You need **maximum accuracy** (full-precision models are better for this).  \n❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).  \n\n---\n\n### **Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)**  \nThese models are optimized for **extreme memory efficiency**, making them ideal for **low-power devices** or **large-scale deployments** where memory is a critical constraint.  \n\n- **IQ3_XS**: Ultra-low-bit quantization (3-bit) with **extreme memory efficiency**.  \n  - **Use case**: Best for **ultra-low-memory devices** where even Q4_K is too large.  \n  - **Trade-off**: Lower accuracy compared to higher-bit quantizations.  \n\n- **IQ3_S**: Small block size for **maximum memory efficiency**.  \n  - **Use case**: Best for **low-memory devices** where **IQ3_XS** is too aggressive.  \n\n- **IQ3_M**: Medium block size for better accuracy than **IQ3_S**.  \n  - **Use case**: Suitable for **low-memory devices** where **IQ3_S** is too limiting.  \n\n- **Q4_K**: 4-bit quantization with **block-wise optimization** for better accuracy.  \n  - **Use case**: Best for **low-memory devices** where **Q6_K** is too large.  \n\n- **Q4_0**: Pure 4-bit quantization, optimized for **ARM devices**.  \n  - **Use case**: Best for **ARM-based devices** or **low-memory environments**.  \n\n---\n\n### **Summary Table: Model Format Selection**  \n\n| Model Format  | Precision  | Memory Usage  | Device Requirements  | Best Use Case  |  \n|--------------|------------|---------------|----------------------|---------------|  \n| **BF16**     | Highest    | High          | BF16-supported GPU/CPUs  | High-speed inference with reduced memory |  \n| **F16**      | High       | High          | FP16-supported devices | GPU inference when BF16 isn't available |  \n| **Q4_K**     | Medium Low | Low           | CPU or Low-VRAM devices | Best for memory-constrained environments |  \n| **Q6_K**     | Medium     | Moderate      | CPU with more memory | Better accuracy while still being quantized |  \n| **Q8_0**     | High       | Moderate      | CPU or GPU with enough VRAM | Best accuracy among quantized models |  \n| **IQ3_XS**   | Very Low   | Very Low      | Ultra-low-memory devices | Extreme memory efficiency and low accuracy |  \n| **Q4_0**     | Low        | Low           | ARM or low-memory devices | llama.cpp can optimize for ARM devices |  \n\n---\n\n## **Included Files & Details**  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-bf16.gguf`  \n- Model weights preserved in **BF16**.  \n- Use this if you want to **requantize** the model into a different format.  \n- Best if your device supports **BF16 acceleration**.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-f16.gguf`  \n- Model weights stored in **F16**.  \n- Use if your device supports **FP16**, especially if BF16 is not available.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-bf16-q8_0.gguf`  \n- **Output & embeddings** remain in **BF16**.  \n- All other layers quantized to **Q8_0**.  \n- Use if your device supports **BF16** and you want a quantized version.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-f16-q8_0.gguf`  \n- **Output & embeddings** remain in **F16**.  \n- All other layers quantized to **Q8_0**.    \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-q4_k.gguf`  \n- **Output & embeddings** quantized to **Q8_0**.  \n- All other layers quantized to **Q4_K**.  \n- Good for **CPU inference** with limited memory.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-q4_k_s.gguf`  \n- Smallest **Q4_K** variant, using less memory at the cost of accuracy.  \n- Best for **very low-memory setups**.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-q6_k.gguf`  \n- **Output & embeddings** quantized to **Q8_0**.  \n- All other layers quantized to **Q6_K** .  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-q8_0.gguf`  \n- Fully **Q8** quantized model for better accuracy.  \n- Requires **more memory** but offers higher precision.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-iq3_xs.gguf`  \n- **IQ3_XS** quantization, optimized for **extreme memory efficiency**.  \n- Best for **ultra-low-memory devices**.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-iq3_m.gguf`  \n- **IQ3_M** quantization, offering a **medium block size** for better accuracy.  \n- Suitable for **low-memory devices**.  \n\n### `Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF-q4_0.gguf`  \n- Pure **Q4_0** quantization, optimized for **ARM devices**.  \n- Best for **low-memory environments**.\n- Prefer IQ4_NL for better accuracy.\n\n# <span id=\"testllm\" style=\"color: #7F7FFF;\">🚀 If you find these models useful</span>\n\nPlease click like ❤ . Also I'd really appreciate it if you could test my Network Monitor Assistant at 👉 [Network Monitor Assitant](https://readyforquantum.com).\n\n💬 Click the **chat icon** (bottom right of the main and dashboard pages) . Choose a LLM; toggle between the LLM Types TurboLLM -> FreeLLM -> TestLLM.\n\n### What I'm Testing\n\nI'm experimenting with **function calling** against my network monitoring service. Using small open source models. I am into the question \"How small can it go and still function\".\n\n🟡 **TestLLM** – Runs the current testing model using llama.cpp on 6 threads of a Cpu VM (Should take about 15s to load. Inference speed is quite slow and it only processes one user prompt at a time—still working on scaling!). If you're curious, I'd be happy to share how it works! .\n\n### The other Available AI Assistants\n\n🟢 **TurboLLM** – Uses **gpt-4o-mini** Fast! . Note: tokens are limited since OpenAI models are pricey, but you can [Login](https://readyforquantum.com) or [Download](https://readyforquantum.com/download/?utm_source=huggingface&utm_medium=referral&utm_campaign=huggingface_repo_readme) the Quantum Network Monitor agent to get more tokens, Alternatively use the TestLLM .\n\n🔵 **HugLLM** – Runs **open-source Hugging Face models** Fast, Runs small models (≈8B) hence lower quality, Get 2x more tokens (subject to Hugging Face API availability)\n\n### Final Word\n\nI fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is [open source](https://github.com/Mungert69). Feel free to use whatever you find helpful.\n\nIf you appreciate the work, please consider [buying me a coffee](https://www.buymeacoffee.com/mahadeva) ☕. Your support helps cover service costs and allows me to raise token limits for everyone.\n\nI'm also open to job opportunities or sponsorship.\n\nThank you! 😊\n\n\n\n# Llama-3.3-Nemotron-Super-49B-v1\n\n## Model Overview \n\nLlama-3.3-Nemotron-Super-49B-v1 is a large language model (LLM) which is a derivative of [Meta Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) (AKA the *reference model*). It is a reasoning model that is post trained for reasoning, human chat preferences, and tasks, such as RAG and tool calling. The model supports a context length of 128K tokens.\n\nLlama-3.3-Nemotron-Super-49B-v1 is a model which offers a great tradeoff between model accuracy and efficiency. Efficiency (throughput) directly translates to savings. Using a novel Neural Architecture Search (NAS) approach, we greatly reduce the model’s memory footprint, enabling larger workloads, as well as fitting the model on a single GPU at high workloads (H200). This NAS approach enables the selection of a desired point in the accuracy-efficiency tradeoff. For more information on the NAS approach, please refer to [this paper](https://arxiv.org/abs/2411.19146).\n\nThe model underwent a multi-phase post-training process to enhance both its reasoning and non-reasoning capabilities. This includes a supervised fine-tuning stage for Math, Code, Reasoning, and Tool Calling as well as multiple reinforcement learning (RL) stages using REINFORCE (RLOO) and Online Reward-aware Preference Optimization (RPO) algorithms for both chat and instruction-following. The final model checkpoint is obtained after merging the final SFT and Online RPO checkpoints. For more details on how the model was trained, please see [this blog](https://developer.nvidia.com/blog/build-enterprise-ai-agents-with-advanced-open-nvidia-llama-nemotron-reasoning-models/).  \n![Training Process](flow.png)\n\nThis model is part of the Llama Nemotron Collection. You can find the other model(s) in this family here:\n- [Llama-3.1-Nemotron-Nano-8B-v1](https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-8B-v1)\n\nThis model is ready for commercial use. \n\n## License/Terms of Use\n\nGOVERNING TERMS: Your use of this model is governed by the [NVIDIA Open Model License.](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) \\\nAdditional Information: [Llama 3.3 Community License Agreement](https://www.llama.com/llama3_3/license/). Built with Llama.\n\n**Model Developer:** NVIDIA\n\n**Model Dates:** Trained between November 2024 and February 2025\n\n**Data Freshness:**  The pretraining data has a cutoff of 2023 per Meta Llama 3.3 70B\n\n### Use Case: <br>\nDevelopers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks. <br>\n\n### Release Date:  <br>\n3/18/2025 <br>\n\n## References\n* [[2411.19146] Puzzle: Distillation-Based NAS for Inference-Optimized LLMs](https://arxiv.org/abs/2411.19146)\n* [[2502.00203] Reward-aware Preference Optimization: A Unified Mathematical Framework for Model Alignment](https://arxiv.org/abs/2502.00203)\n\n## Model Architecture\n**Architecture Type:** Dense decoder-only Transformer model \\\n**Network Architecture:** Llama 3.3 70B Instruct, customized through Neural Architecture Search (NAS)\n\nThe model is a derivative of Meta’s Llama-3.3-70B-Instruct, using Neural Architecture Search (NAS). The NAS algorithm results in non-standard and non-repetitive blocks. This includes the following: \n* Skip attention: In some blocks, the attention is skipped entirely, or replaced with a single linear layer.  \n* Variable FFN: The expansion/compression ratio in the FFN layer is different between blocks.\n\nWe utilize a block-wise distillation of the reference model, where for each block we create multiple variants providing different tradeoffs of quality vs. computational complexity, discussed in more depth below. We then search over the blocks to create a model which meets the required throughput and memory (optimized for a single H100-80GB GPU) while minimizing the quality degradation. The model then undergoes knowledge distillation (KD), with a focus on English single and multi-turn chat use-cases. The KD step included 40 billion tokens consisting of a mixture of 3 datasets - FineWeb, Buzz-V1.2 and Dolma.\n\n## Intended use\n\nLlama-3.3-Nemotron-Super-49B-v1 is a general purpose reasoning and chat model intended to be used in English and coding languages. Other non-English languages (German, French, Italian, Portuguese, Hindi, Spanish, and Thai) are also supported. \n\n## Input\n- **Input Type:** Text\n- **Input Format:** String\n- **Input Parameters:** One-Dimensional (1D)\n- **Other Properties Related to Input:** Context length up to 131,072 tokens\n\n## Output\n- **Output Type:** Text\n- **Output Format:** String\n- **Output Parameters:** One-Dimensional (1D)\n- **Other Properties Related to Output:** Context length up to 131,072 tokens\n\n## Model Version\n1.0 (3/18/2025)\n\n## Software Integration\n- **Runtime Engine:** Transformers\n- **Recommended Hardware Microarchitecture Compatibility:** \n  - NVIDIA Hopper\n  - NVIDIA Ampere\n\n## Quick Start and Usage Recommendations:\n\n1. Reasoning mode (ON/OFF) is controlled via the system prompt, which must be set as shown in the example below. All instructions should be contained within the user prompt\n2. We recommend setting temperature to `0.6`, and Top P to `0.95` for Reasoning ON mode\n3. We recommend using greedy decoding for Reasoning OFF mode\n4. We have provided a list of prompts to use for evaluation for each benchmark where a specific template is required\n\nYou can try this model out through the preview API, using this link: [Llama-3_3-Nemotron-Super-49B-v1](https://build.nvidia.com/nvidia/llama-3_3-nemotron-super-49b-v1).\n\nSee the snippet below for usage with [Hugging Face Transformers](https://huggingface.co/docs/transformers/main/en/index) library. Reasoning mode (ON/OFF) is controlled via system prompt. Please see the example below\n\nWe recommend using the *transformers* package with version 4.48.3.   \nExample of reasoning on:\n\n```py\nimport torch\nimport transformers\n\nmodel_id = \"nvidia/Llama-3_3-Nemotron-Super-49B-v1\"\nmodel_kwargs = {\"torch_dtype\": torch.bfloat16, \"trust_remote_code\": True, \"device_map\": \"auto\"}\ntokenizer = transformers.AutoTokenizer.from_pretrained(model_id)\ntokenizer.pad_token_id = tokenizer.eos_token_id\n\npipeline = transformers.pipeline(\n   \"text-generation\",\n   model=model_id,\n   tokenizer=tokenizer,\n   max_new_tokens=32768,\n   temperature=0.6,\n   top_p=0.95,\n   **model_kwargs\n)\n\nthinking = \"on\"\n\nprint(pipeline([{\"role\": \"system\", \"content\": f\"detailed thinking {thinking}\"},{\"role\": \"user\", \"content\": \"Solve x*(sin(x)+2)=0\"}]))\n```\n\nExample of reasoning off:\n\n```py\nimport torch\nimport transformers\n\nmodel_id = \"nvidia/Llama-3_3-Nemotron-Super-49B-v1\"\nmodel_kwargs = {\"torch_dtype\": torch.bfloat16, \"trust_remote_code\": True, \"device_map\": \"auto\"}\ntokenizer = transformers.AutoTokenizer.from_pretrained(model_id)\ntokenizer.pad_token_id = tokenizer.eos_token_id\n\npipeline = transformers.pipeline(\n   \"text-generation\",\n   model=model_id,\n   tokenizer=tokenizer,\n   max_new_tokens=32768,\n   do_sample=False,\n   **model_kwargs\n)\n\n# Thinking can be \"on\" or \"off\"\nthinking = \"off\"\n\nprint(pipeline([{\"role\": \"system\", \"content\": f\"detailed thinking {thinking}\"},{\"role\": \"user\", \"content\": \"Solve x*(sin(x)+2)=0\"}]))\n```\n\n## Inference:\n\n**Engine:**\n  - Transformers \n\n**Test Hardware:** \n  - FP8: 1x NVIDIA H100-80GB GPU (Coming Soon!)\n  - BF16: \n    - 2x NVIDIA H100-80GB \n    - 2x NVIDIA A100-80GB GPUs\n      \n**[Preferred/Supported] Operating System(s):** Linux <br>\n\n## Training Datasets\n\nA large variety of training data was used for the knowledge distillation phase before post-training pipeline, 3 of which included: FineWeb, Buzz-V1.2, and Dolma.\n\nThe data for the multi-stage post-training phases for improvements in Code, Math, and Reasoning is a compilation of SFT and RL data that supports improvements of math, code, general reasoning, and instruction following capabilities of the original Llama instruct model. \n\nIn conjunction with this model release, NVIDIA has released 30M samples of post-training data, as public and permissive. Please see [Llama-Nemotron-Postraining-Dataset-v1](https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset-v1).\n\nDistribution of the domains is as follows:\n\n| Category | Value     |\n|----------|-----------|\n| math     | 19,840,970|\n| code     | 9,612,677 |\n| science     | 708,920    |\n| instruction following       | 56,339    |\n| chat     | 39,792    |\n| safety   | 31,426    |\n\nPrompts have been sourced from either public and open corpus or synthetically generated. Responses were synthetically generated by a variety of models, with some prompts containing responses for both reasoning on and off modes, to train the model to distinguish between two modes. \n\n\n**Data Collection for Training Datasets:**\n\n- Hybrid: Automated, Human, Synthetic\n\n**Data Labeling for Training Datasets:**\n\n- Hybrid: Automated, Human, Synthetic\n\n## Evaluation Datasets \n\nWe used the datasets listed below to evaluate Llama-3.3-Nemotron-Super-49B-v1. \n\nData Collection for Evaluation Datasets:\n\n- Hybrid: Human/Synthetic\n\nData Labeling for Evaluation Datasets:\n\n- Hybrid: Human/Synthetic/Automatic\n\n## Evaluation Results  \nThese results contain both “Reasoning On”, and “Reasoning Off”. We recommend using temperature=`0.6`, top_p=`0.95` for “Reasoning On” mode, and greedy decoding for “Reasoning Off” mode. All evaluations are done with 32k sequence length. We run the benchmarks up to 16 times and average the scores to be more accurate.\n\n> NOTE: Where applicable, a Prompt Template will be provided. While completing benchmarks, please ensure that you are parsing for the correct output format as per the provided prompt in order to reproduce the benchmarks seen below. \n\n### Arena-Hard\n\n| Reasoning Mode | Score |  \n|--------------|------------|  \n| Reasoning Off | 88.3 | \n\n### MATH500\n\n| Reasoning Mode | pass@1 |  \n|--------------|------------|  \n| Reasoning Off | 74.0 |   \n| Reasoning On | 96.6  |\n\nUser Prompt Template: \n\n```\n\"Below is a math question. I want you to reason through the steps and then give a final answer. Your final answer should be in \\boxed{}.\\nQuestion: {question}\"\n```\n### AIME25\n\n| Reasoning Mode | pass@1 |  \n|--------------|------------|  \n| Reasoning Off | 13.33 |   \n| Reasoning On | 58.4 |\n\nUser Prompt Template: \n\n```\n\"Below is a math question. I want you to reason through the steps and then give a final answer. Your final answer should be in \\boxed{}.\\nQuestion: {question}\"\n```\n\n### GPQA\n\n| Reasoning Mode | pass@1 |  \n|--------------|------------|  \n| Reasoning Off | 50 |   \n| Reasoning On | 66.67 |\n\nUser Prompt Template: \n\n```\n\"What is the correct answer to this question: {question}\\nChoices:\\nA. {option_A}\\nB. {option_B}\\nC. {option_C}\\nD. {option_D}\\nLet's think step by step, and put the final answer (should be a single letter A, B, C, or D) into a \\boxed{}\"\n```\n\n### IFEval\n\n| Reasoning Mode | Strict:Instruction |  \n|--------------|------------|  \n| Reasoning Off | 89.21 | \n\n### BFCL V2 Live\n\n| Reasoning Mode | Score |  \n|--------------|------------|  \n| Reasoning Off | 73.7 | \n\nUser Prompt Template:\n\n```\nYou are an expert in composing functions. You are given a question and a set of possible functions. \nBased on the question, you will need to make one or more function/tool calls to achieve the purpose. \nIf none of the function can be used, point it out. If the given question lacks the parameters required by the function,\nalso point it out. You should only return the function call in tools call sections.\n\nIf you decide to invoke any of the function(s), you MUST put it in the format of <TOOLCALL>[func_name1(params_name1=params_value1, params_name2=params_value2...), func_name2(params)]</TOOLCALL>\n\nYou SHOULD NOT include any other text in the response.\nHere is a list of functions in JSON format that you can invoke.\n\n<AVAILABLE_TOOLS>{functions}</AVAILABLE_TOOLS>\n\n{user_prompt}\n```\n\n### MBPP 0-shot\n\n| Reasoning Mode | pass@1 |  \n|--------------|------------|  \n| Reasoning Off | 84.9|   \n| Reasoning On | 91.3 |\n\nUser Prompt Template:\n\n````\nYou are an exceptionally intelligent coding assistant that consistently delivers accurate and reliable responses to user instructions.\n\n@@ Instruction\nHere is the given problem and test examples:\n{prompt}\nPlease use the python programming language to solve this problem.\nPlease make sure that your code includes the functions from the test samples and that the input and output formats of these functions match the test samples.\nPlease return all completed codes in one code block.\nThis code block should be in the following format:\n```python\n# Your codes here\n```\n````\n\n### MT-Bench\n\n| Reasoning Mode | Score |  \n|--------------|------------|  \n| Reasoning Off | 9.17 |\n\n## Ethical Considerations:\n\nNVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications.  When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. \n\nFor more detailed information on ethical considerations for this model, please see the Model Card++ [Explainability](explainability.md), [Bias](bias.md), [Safety & Security](safety.md), and [Privacy](privacy.md) Subcards.  \n\nPlease report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).\n",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "nvidia",
    "llama-3",
    "pytorch",
    "text-generation",
    "en",
    "arxiv:2411.19146",
    "arxiv:2502.00203",
    "license:other",
    "endpoints_compatible",
    "region:us",
    "imatrix",
    "conversational"
  ],
  "likes": 5,
  "downloads": 98,
  "gated": false,
  "private": false,
  "last_modified": "2025-09-24T15:41:43.000Z",
  "created_at": "2025-03-29T03:22:36.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "transformers"
}

{
  "_id": "67e7677c2c2aead26d89986d",
  "id": "Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF",
  "modelId": "Mungert/Llama-3_3-Nemotron-Super-49B-v1-GGUF",
  "sha": "2ced76904238df46498568fb64c1df822dacddb7",
  "createdAt": "2025-03-29T03:22:36.000Z",
  "lastModified": "2025-09-24T15:41:43.000Z",
  "author": "Mungert",
  "downloads": 98,
  "likes": 5,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "transformers",
  "siblings_count": 41
}