MiniMax-M2.7-IQ4_XS-HQ-GGUF

MiniMax-M2.7-IQ4_XS-HQ-GGUF

Just testing out a new recipe on MiniMax-M2.7

This repository now contains the current HQ v2 release. Earlier uploaded variants were removed after an issue was discovered during evaluation.

Quantization

• Converted from the original FP8 safetensors weights to a BF16 GGUF, then quantized to IQ4_XS
• Mixed-tensor overrides for higher quality

What is protected

Compared with a flat quant, this build prioritizes higher precision for tensor groups that are more important for output stability and quality, such as:

• LM head
• token embeddings
• norm tensors
• selected routing / control tensors
• selected attention-side tensors

Usage

Load the first shard in llama.cpp with --jinja --chat-template-file /path/to/minimax-m2.7-fixed-llamacpp-template.jinja; the remaining shards will be detected automatically.

Notes

• Format: GGUF
• Intended runtime: llama.cpp and compatible loaders
• Quant family: IQ4_XS
• Focus: quality-preserving compact quantization
• Current evaluation status: perplexity working (no NaNs!), KLD pending