---

base_model: LiquidAI/LFM2.5-24B-A2B

library_name: gguf

pipeline_tag: text-generation

tags:

- gguf

- llama.cpp

- quantized

---

LFM2.5-24B-A2B GGUF

GGUF quantizations of LiquidAI/LFM2.5-24B-A2B, covering 31 files (365.9 GB total).

Files

| File | Quant | Size | Notes |

|---|---|---:|---|

| LFM2.5-24B-A2B-BF16.gguf | BF16 | 44.67 GB | Full precision source. Every quant below is cut from this file. |

| LFM2.5-24B-A2B-Q8_0.gguf | Q8_0 | 23.74 GB | Effectively lossless. Use when disk and RAM are not the constraint. |

| LFM2.5-24B-A2B-Q6_K.gguf | Q6_K | 18.34 GB | Near-lossless; the last stop before quality becomes measurable. |

| LFM2.5-24B-A2B-Q5_K_M.gguf | Q5_K_M | 15.86 GB | Very good quality, noticeably smaller than Q6_K. |

| LFM2.5-24B-A2B-Q5_K_S.gguf | Q5_K_S | 15.41 GB | Slightly smaller than Q5_K_M for a slight quality cost. |

| LFM2.5-24B-A2B-Q5_1.gguf | Q5_1 | 16.78 GB | Legacy. Prefer Q5_K_M. |

| LFM2.5-24B-A2B-Q5_0.gguf | Q5_0 | 15.41 GB | Legacy. Prefer Q5_K_M. |

| LFM2.5-24B-A2B-Q4_K_M.gguf | Q4_K_M | 13.53 GB | The usual default. Best quality-per-byte for most people. |

| LFM2.5-24B-A2B-Q4_K_S.gguf | Q4_K_S | 12.72 GB | A little smaller than Q4_K_M, a little worse. |

| LFM2.5-24B-A2B-IQ4_NL.gguf | IQ4_NL | 12.65 GB | Non-linear 4-bit; good on hardware without fast K-quant kernels. |

| LFM2.5-24B-A2B-IQ4_XS.gguf | IQ4_XS | 11.96 GB | Best sub-4.5bpw option; usually beats Q4_K_S at a smaller size. |

| LFM2.5-24B-A2B-Q4_1.gguf | Q4_1 | 14.03 GB | Legacy. Prefer Q4_K_M. |

| LFM2.5-24B-A2B-Q4_0.gguf | Q4_0 | 12.65 GB | Legacy round-to-nearest. Prefer Q4_K_M unless a runtime needs this. |

| LFM2.5-24B-A2B-MXFP4_MOE.gguf | MXFP4_MOE | 12.39 GB | MoE-only 4-bit microscaling format for the expert tensors. |

| LFM2.5-24B-A2B-Q3_K_L.gguf | Q3_K_L | 11.57 GB | Small, with real quality loss. Usable when RAM is tight. |

| LFM2.5-24B-A2B-Q3_K_M.gguf | Q3_K_M | 10.68 GB | Smaller again; noticeable degradation. |

| LFM2.5-24B-A2B-IQ3_M.gguf | IQ3_M | 9.80 GB | Strong at ~3.7bpw, clearly better than Q3_K_M. |

| LFM2.5-24B-A2B-IQ3_S.gguf | IQ3_S | 9.71 GB | Slightly smaller than IQ3_M. |

| LFM2.5-24B-A2B-Q3_K_S.gguf | Q3_K_S | 9.71 GB | Aggressive. Prefer IQ3_M at a similar size. |

| LFM2.5-24B-A2B-IQ3_XS.gguf | IQ3_XS | 9.18 GB | Aggressive but coherent. |

| LFM2.5-24B-A2B-IQ3_XXS.gguf | IQ3_XXS | 8.65 GB | Very aggressive; the last coherent step down. |

| LFM2.5-24B-A2B-Q2_K.gguf | Q2_K | 8.20 GB | Very small, heavily degraded. For experimentation. |

| LFM2.5-24B-A2B-IQ2_M.gguf | IQ2_M | 2.02 GB | The smallest size most people find usable. |

| LFM2.5-24B-A2B-Q2_K_S.gguf | Q2_K_S | 7.61 GB | Smaller than Q2_K, at a further quality cost. |

| LFM2.5-24B-A2B-IQ2_S.gguf | IQ2_S | 6.67 GB | Below the usual usability line. |

| LFM2.5-24B-A2B-IQ2_XS.gguf | IQ2_XS | 6.60 GB | Experimental. |

| LFM2.5-24B-A2B-IQ2_XXS.gguf | IQ2_XXS | 5.93 GB | Experimental. |

| LFM2.5-24B-A2B-Q2_0.gguf | Q2_0 | 6.44 GB | Extreme, group-64. Included for completeness. |

| LFM2.5-24B-A2B-IQ1_M.gguf | IQ1_M | 5.10 GB | Extreme. Expect substantial degradation. |

| LFM2.5-24B-A2B-IQ1_S.gguf | IQ1_S | 4.59 GB | Extreme. Expect substantial degradation. |

| LFM2.5-24B-A2B-Q1_0.gguf | Q1_0 | 3.33 GB | Extreme. Included for completeness. |

Which one should I download?

Pick the largest file that leaves a couple of gigabytes of headroom on the device you will run it on — the model has to fit in RAM (or VRAM, if you are offloading) alongside the KV cache and the OS.

• Plenty of memory: Q6_K or Q8_0.
• The usual choice: Q4_K_M.
• Tight on memory: IQ4_XS, then IQ3_M, then IQ2_M.
• The IQ files generally beat a Q file of similar size, at the cost of slightly slower inference on some hardware.

Usage

llama-cli -hf NANI-Nithin/LFM2.5-24B-A2B-GGUF:Q4_K_M -p "Hello"

Or download one file and point at it directly:

huggingface-cli download NANI-Nithin/LFM2.5-24B-A2B-GGUF LFM2.5-24B-A2B-Q4_K_M.gguf --local-dir .
llama-cli -m LFM2.5-24B-A2B-Q4_K_M.gguf -p "Hello"

---

Quantized with llama.cpp by AgentQuantix on 2026-10-14. Licensing follows the base model.