ox-ox/minimax-m2.7-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "other",
    "license_name": "modified-mit",
    "base_model": "MiniMaxAI/MiniMax-M2.7",
    "tags": [
      "llama-cpp",
      "gguf",
      "quantized",
      "apple-silicon",
      "mixture-of-experts",
      "minimax",
      "reasoning"
    ],
    "language": [
      "en",
      "fr"
    ],
    "pipeline_tag": "text-generation",
    "frontmatter": {
      "license": "other",
      "license_name": "modified-mit",
      "base_model": "MiniMaxAI/MiniMax-M2.7",
      "tags": [
        "llama-cpp",
        "gguf",
        "quantized",
        "apple-silicon",
        "mixture-of-experts",
        "minimax",
        "reasoning"
      ],
      "language": [
        "en",
        "fr"
      ],
      "pipeline_tag": "text-generation"
    },
    "hero_image_url": "",
    "summary": "High-precision GGUF quants of MiniMax-M2.7 (229B parameters) Mixture of Experts model. Optimized for local inference on high-RAM setups, particularly Apple Silicon (M3 Max/Ultra).",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: other\nlicense_name: modified-mit\nbase_model: MiniMaxAI/MiniMax-M2.7\ntags:\n- llama-cpp\n- gguf\n- quantized\n- apple-silicon\n- mixture-of-experts\n- minimax\n- reasoning\nlanguage:\n- en\n- fr\npipeline_tag: text-generation\n---\n\n# MiniMax-M2.7-GGUF (229B MoE)\n\nHigh-precision GGUF quants of [MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) (229B parameters) Mixture of Experts model. Optimized for local inference on high-RAM setups, particularly Apple Silicon (M3 Max/Ultra).\n\n## Perplexity Validation (WikiText-2)\n\n| Quant | PPL (c=512, seed=1337) | Speed (M3 Max 128GB) |\n|-------|------------------------|----------------------|\n| Q3_K_L | **8.4400 ± 0.065** | **28.52 t/s** |\n\n*Baseline — MiniMax-M2.5 Q3_K_L: 8.7948 PPL, 28.7 t/s*\n\n## Available Quants\n\n| File | Method | Size | Use Case |\n|------|--------|------|----------|\n| `minimax-m2.7-Q3_K_L.gguf` | Q3_K_L | ~110 GB | Sweet spot for 128GB Macs. Runs natively in RAM. |\n| `minimax-m2.7-Q8_0.gguf` | Q8_0 | ~243 GB | Maximum precision. Requires 256GB+ unified memory. |\n\n## Model Highlights\n\n- **Self-evolution:** M2.7 participated in its own training — autonomously optimized a programming scaffold over 100+ rounds, achieving 30% performance improvement\n- **MLE Bench Lite:** 66.6% medal rate (22 ML competitions), second only to Opus-4.6 and GPT-5.4\n- **SWE-Pro:** 56.22% — matches GPT-5.3-Codex\n- **SWE Multilingual:** 76.5 | **Multi SWE Bench:** 52.7\n- **VIBE-Pro:** 55.6% — nearly on par with Opus 4.6\n- **Terminal Bench 2:** 57.0% | **NL2Repo:** 39.8%\n- **GDPval-AA ELO:** 1495 — highest among open-source models\n- **Native Agent Teams** support for multi-agent collaboration\n\n## Model Details\n\n- **Architecture:** MiniMax-M2 (Mixture of Experts) with 256 experts (8 active per token)\n- **Parameters:** ~229B total\n- **Quantization Process:** FP8 safetensors → Q8_0 → Q3_K_L via llama.cpp\n- **Context Window:** Up to 196k tokens\n- **Chat Template:** Includes official Jinja template for `<think>` tag handling\n\n## Recommended Inference Parameters\n\n```\ntemperature=1.0, top_p=0.95, top_k=40\n```\n\nDefault system prompt:\n```\nYou are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax.\n```\n\n## Usage\n\n### 1. Install llama.cpp\n\n```bash\ngit clone https://github.com/ggml-org/llama.cpp\ncd llama.cpp\ncmake -B build -DGGML_METAL=ON\ncmake --build build --config Release -j\n```\n\n### 2. Download the model\n\n```bash\n# Q3_K_L (128GB Mac)\nhuggingface-cli download ox-ox/MiniMax-M2.7-GGUF \\\n  minimax-m2.7-Q3_K_L.gguf --local-dir .\n\n# Q8_0 (256GB+)\nhuggingface-cli download ox-ox/MiniMax-M2.7-GGUF \\\n  minimax-m2.7-Q8_0.gguf --local-dir .\n```\n\n### 3. Unlock Metal memory limit (128GB Mac only)\n\nThe model weights use ~118GB. Run this before launching to allow full GPU offload:\n\n```bash\nsudo sysctl iogpu.wired_limit_mb=122000\n```\n\n### 4. Run\n\n```bash\n./build/bin/llama-server -m minimax-m2.7-Q3_K_L.gguf \\\n  -ngl 99 \\\n  --ctx-size 512 \\\n  -b 512 -ub 512 \\\n  --port 8080 \\\n  --jinja\n```\n\n> ⚠️ **License:** Non-commercial use only. Commercial use requires written authorization from MiniMax. See [LICENSE](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE).",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "llama-cpp",
    "quantized",
    "apple-silicon",
    "mixture-of-experts",
    "minimax",
    "reasoning",
    "text-generation",
    "en",
    "fr",
    "base_model:MiniMaxAI/MiniMax-M2.7",
    "base_model:quantized:MiniMaxAI/MiniMax-M2.7",
    "license:other",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 1,
  "downloads": 510,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-12T11:31:28.000Z",
  "created_at": "2026-04-12T01:55:03.000Z",
  "pipeline_tag": "text-generation",
  "library_name": ""
}

{
  "_id": "69dafb77e48aad1702e810ad",
  "id": "ox-ox/MiniMax-M2.7-GGUF",
  "modelId": "ox-ox/MiniMax-M2.7-GGUF",
  "sha": "2b93fc0d4e8354ca870eda5716d1800417ae9a16",
  "createdAt": "2026-04-12T01:55:03.000Z",
  "lastModified": "2026-04-12T11:31:28.000Z",
  "author": "ox-ox",
  "downloads": 510,
  "likes": 1,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "",
  "siblings_count": 4
}