The canonical BF16 Transformers checkpoint is available at Unmid/Mew1-2.6B. Additional importance-matrix quantizations are available from mradermacher.

Provided files

Files are sorted from smallest to largest. Sizes are the actual decimal file sizes reported by the Hugging Face Hub.

File	Quant	Size	Notes
Mew1-2.6B.Q2_K.gguf	Q2_K	1.09 GB	Smallest; substantial quality loss
Mew1-2.6B.Q4_0.gguf	Q4_0	1.59 GB	Fast legacy 4-bit quant
Mew1-2.6B.Q4_K_M.gguf	Q4_K_M	1.67 GB	Recommended size/quality balance
Mew1-2.6B.Q5_K_M.gguf	Q5_K_M	1.94 GB	Higher quality
Mew1-2.6B.Q6_K.gguf	Q6_K	2.22 GB	Very high quality
Mew1-2.6B.Q8_0.gguf	Q8_0	2.87 GB	Largest quantized version

Download and run

Install the Hugging Face CLI and download the recommended quant:

bash

pip install -U huggingface_hub
hf download Unmid/Mew1-2.6B-GGUF Mew1-2.6B.Q4_K_M.gguf --local-dir .

Run it with llama.cpp:

bash

llama-cli \
  -m Mew1-2.6B.Q4_K_M.gguf \
  -c 4096 \
  --temp 0.7 \
  --top-k 50 \
  --top-p 0.9 \
  --repeat-penalty 1.1 \
  -cnv

The model metadata supports a context of approximately 128K tokens, but memory use grows with context length. Start with 4K–16K and increase only if your hardware allows it.

Quant selection

Q4_K_M — recommended starting point for most users.
Q5_K_M or Q6_K — use when quality matters more than memory.
Q8_0 — largest quantized file; useful when memory is plentiful.
Q2_K — only for severe memory constraints.

LEAP manifests

The leap/ directory contains one manifest per quantization. Each manifest references the corresponding root-level GGUF filename and uses the checkpoint chat template.

Model lineage

LiquidAI/LFM2.5-2.6B → SC117/LFM2.5-2.6B-Uncensored → Unmid/Mew1-2.6B → this GGUF release.

This is a community derivative and is not an official Liquid AI or SC117 release.

License

Distributed under the LFM Open License v1.0. Review LICENSE, including its attribution and commercial-use conditions.