johnbean393/LFM2.5-8B-A1B-Base-GGUF

LFM2.5-8B-A1B-Base GGUF

This repository contains an imatrix-assisted Q4_K_M GGUF quantization of LiquidAI/LFM2.5-8B-A1B-Base.

Files

• LFM2.5-8B-A1B-Base-Q4_K_M.gguf: quantized GGUF model
• LFM2.5-8B-A1B-Base-imatrix.gguf: importance matrix used during quantization

Build Notes

The source safetensors file was verified against the upstream SHA-256:

9a57e48f4f70c56d4a2c7718de5cd89bee5c0402ee74b4e5da3ec519192b5c77

The conversion used llama.cpp. The local converter needed the LFM2.5 tokenizer pre-tokenizer hash mapped to llama.cpp's existing lfm2 tokenizer identifier.

The imatrix calibration run completed, with partial coverage reported for a few MoE expert tensors in block 9. This can occur when a short calibration corpus does not route tokens through every expert equally.

Smoke Test

Before publishing, the quantized GGUF was loaded with llama.cpp in single-turn conversation mode and checked on arithmetic, JSON formatting, and a simple Python function prompt.