LFM2.5-8B-A1B-GGUF

Liquid LFM2.5-8B-A1B GGUF — Quantized by BatiAI

<p align="center"> <a href="https://flow.bati.ai"><img src="https://img.shields.io/badge/BatiFlow-macOS%20AI%20Automation-blue?style=for-the-badge&logo=apple" alt="BatiFlow"></a> <a href="https://ollama.com/batiai/lfm2.5-8b"><img src="https://img.shields.io/badge/Ollama-batiai%2Flfm2.5--8b-green?style=for-the-badge" alt="Ollama"></a> </p>

Quantizations of Liquid AI LFM2.5-8B-A1B (8.3B total / 1.5B active MoE, reasoning + tool calling) for on-device AI on Mac. Built and verified by BatiAI for BatiFlow.

Why LFM2.5-8B-A1B?

• 8.3B total, only 1.5B active — MoE (32 experts, 4 active) gives 8B-class quality at ~1.5B-class speed
• Hybrid architecture — 24 layers: 18 double-gated LIV convolution + 6 GQA attention. Built for edge, not GPUs
• Reasoning model — explicit chain-of-thought before the final answer
• Native tool calling — Pythonic or JSON function calls (<|tool_call_start|> / <|tool_call_end|>)
• 128K context, 38T-token training budget, 128K vocab (better non-Latin tokenization incl. Korean)
• Blazing on-device — Liquid reports 253 tok/s on M5 Max, 146 tok/s on Ryzen AI Max+ 395, all under 6GB
• Released May 28, 2026

Quick Start

ollama pull batiai/lfm2.5-8b:q4

Available Quantizations

Mac note on Q3_K_M: in every model we've benchmarked on Apple Silicon, Q3_K_M generated slower than Q4_K_M despite the smaller file — Granite 4.1 (+27%), Gemma 4 26B (+12%), Qwen3.8‑27B (+18%), Qwen3.6‑27B (+8%), on both M4 Max and M4 mini. Metal's Q3_K path is limited by dequantization compute rather than bandwidth. We have not measured this particular model's Q3/Q4 pair yet, so treat it as a strong prior, not a measurement: if Q4_K_M fits, take it. On CUDA the two are effectively tied, so this applies to Macs only.

Lower quants (IQ3_XXS, Q2_K_S) and imatrix variants are not provided in the official Liquid GGUF repo — BatiAI adds them for 8GB Macs and tighter footprints.

RAM Requirements

This model is tiny in active params — even the largest quant fits comfortably on 16GB. With only 1.5B active params, throughput stays high regardless of quant.

Why BatiAI Quantization?

Technical Details

• Original Model: LiquidAI/LFM2.5-8B-A1B
• Architecture: lfm2_moe hybrid — 18 LIV conv + 6 GQA layers, 32 experts / 4 active per token
• Parameters: 8.3B total, 1.5B active
• Context: 131,072 tokens
• Training: 38T tokens
• License: LFM Open License v1.0 (Apache-2.0 based; free commercial use under $10M annual revenue)
• Quantized with: llama.cpp + imatrix
• Quantized by: BatiAI

About BatiFlow

BatiFlow — free, on-device AI automation for Mac. 5MB app, 100% local, unlimited.

License

Quantized from LiquidAI/LFM2.5-8B-A1B. License: LFM Open License v1.0 — see https://www.liquid.ai/lfm-license. Original Liquid AI attribution retained. Free for commercial use by organizations under $10M USD annual revenue; above that threshold, contact Liquid AI.

Benchmarks

<!-- BENCH-START -->

Benchmarks pending — measured on target Mac hardware via bench.sh.

<!-- BENCH-END -->