---
license: mit
base_model: MiniMaxAI/MiniMax-M2.5
tags:
- gguf
- moe
- minimax
- llama.cpp
- applesilicon
- reasoning
- conversational
model_creator: MiniMaxAI
pipeline_tag: text-generation
---

# MiniMax-M2.5-GGUF (230B MoE)

High-precision GGUF quants of the **MiniMax-M2.5** (230B parameters) Mixture of Experts model. These versions are specifically optimized for local inference on high-RAM setups, particularly **Apple Silicon (M3 Max/Ultra)**.

🔬 Perplexity Validation (WikiText-2):

Final PPL: 8.2213 +/- 0.09

Context: 4096 / 32 chunks

Outcome: The Q3_K_L quantization maintains high logical coherence while boosting speed to 28.7 t/s. Minimal degradation for a ~20GB size reduction vs Q4.
## 🚀 Available Quants

| File Name | Method | Size | Use Case |
| :--- | :--- | :--- | :--- |
| `minimax-m2.5-Q4_K_M.gguf` | Q4_K_M | 138 GB | Highest logic preservation. Requires >128GB RAM or SSD swap. |
| `minimax-m2.5-Q3_K_L.gguf` | **Q3_K_L** | **~110 GB** | **Sweet spot for 128GB Macs**. Runs natively in RAM with high t/s ( 28 ON MAC M3 MAX ). |

## 🛠 Model Details

* **Architecture:** MiniMax-M2 (Mixture of Experts) with 256 experts (8 active per token).
* **Parameters:** ~230B total.
* **Quantization Process:** Unlike automated scripts, these quants were generated from a full **F16 GGUF Master (457GB)** to minimize accumulation of errors during the K-Quant process.
* **Context Window:** Up to 196k tokens (Native support).
* **Chat Template:** Includes the official **Jinja template** for proper handling of interleaved `<think>` tags, separating reasoning from the final response.

## 💻 Usage

Requires **llama.cpp** build 8022 or higher.

### Command Line Example:
```bash
./llama-cli -m minimax-m2.5-Q3_K_L.gguf -n -1 \\
  -c 262000 \\
  -ngl 99 -fa on -ctk q4_0 -ctv q4_0 -b 2048 -ub 1024 --port 8080 --jinja --verbose -sm none --draft 16 -ncmoe 0 --cache-reuse 1024 --draft-p-min 0.5