---

base_model: lightonai/LightOnOCR-3-0.8B

library_name: gguf

pipeline_tag: text-generation

tags:

- gguf

- llama.cpp

- quantized

---

LightOnOCR-3-0.8B GGUF

GGUF quantizations of lightonai/LightOnOCR-3-0.8B, covering 30 files (14.1 GB total).

Files

| File | Quant | Size | Notes |

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

| LightOnOCR-3-0.8B-BF16.gguf | BF16 | 1.41 GB | Full precision source. Every quant below is cut from this file. |

| LightOnOCR-3-0.8B-Q8_0.gguf | Q8_0 | 0.76 GB | Effectively lossless. Use when disk and RAM are not the constraint. |

| LightOnOCR-3-0.8B-Q6_K.gguf | Q6_K | 0.59 GB | Near-lossless; the last stop before quality becomes measurable. |

| LightOnOCR-3-0.8B-Q5_K_M.gguf | Q5_K_M | 0.54 GB | Very good quality, noticeably smaller than Q6_K. |

| LightOnOCR-3-0.8B-Q5_K_S.gguf | Q5_K_S | 0.52 GB | Slightly smaller than Q5_K_M for a slight quality cost. |

| LightOnOCR-3-0.8B-Q5_1.gguf | Q5_1 | 0.55 GB | Legacy. Prefer Q5_K_M. |

| LightOnOCR-3-0.8B-Q5_0.gguf | Q5_0 | 0.52 GB | Legacy. Prefer Q5_K_M. |

| LightOnOCR-3-0.8B-Q4_K_M.gguf | Q4_K_M | 0.49 GB | The usual default. Best quality-per-byte for most people. |

| LightOnOCR-3-0.8B-Q4_K_S.gguf | Q4_K_S | 0.47 GB | A little smaller than Q4_K_M, a little worse. |

| LightOnOCR-3-0.8B-IQ4_NL.gguf | IQ4_NL | 0.48 GB | Non-linear 4-bit; good on hardware without fast K-quant kernels. |

| LightOnOCR-3-0.8B-IQ4_XS.gguf | IQ4_XS | 0.47 GB | Best sub-4.5bpw option; usually beats Q4_K_S at a smaller size. |

| LightOnOCR-3-0.8B-Q4_1.gguf | Q4_1 | 0.50 GB | Legacy. Prefer Q4_K_M. |

| LightOnOCR-3-0.8B-Q4_0.gguf | Q4_0 | 0.47 GB | Legacy round-to-nearest. Prefer Q4_K_M unless a runtime needs this. |

| LightOnOCR-3-0.8B-Q3_K_L.gguf | Q3_K_L | 0.46 GB | Small, with real quality loss. Usable when RAM is tight. |

| LightOnOCR-3-0.8B-Q3_K_M.gguf | Q3_K_M | 0.43 GB | Smaller again; noticeable degradation. |

| LightOnOCR-3-0.8B-IQ3_M.gguf | IQ3_M | 0.42 GB | Strong at ~3.7bpw, clearly better than Q3_K_M. |

| LightOnOCR-3-0.8B-IQ3_S.gguf | IQ3_S | 0.42 GB | Slightly smaller than IQ3_M. |

| LightOnOCR-3-0.8B-Q3_K_S.gguf | Q3_K_S | 0.41 GB | Aggressive. Prefer IQ3_M at a similar size. |

| LightOnOCR-3-0.8B-IQ3_XS.gguf | IQ3_XS | 0.41 GB | Aggressive but coherent. |

| LightOnOCR-3-0.8B-IQ3_XXS.gguf | IQ3_XXS | 0.37 GB | Very aggressive; the last coherent step down. |

| LightOnOCR-3-0.8B-Q2_K.gguf | Q2_K | 0.39 GB | Very small, heavily degraded. For experimentation. |

| LightOnOCR-3-0.8B-IQ2_M.gguf | IQ2_M | 0.35 GB | The smallest size most people find usable. |

| LightOnOCR-3-0.8B-Q2_K_S.gguf | Q2_K_S | 0.39 GB | Smaller than Q2_K, at a further quality cost. |

| LightOnOCR-3-0.8B-IQ2_S.gguf | IQ2_S | 0.34 GB | Below the usual usability line. |

| LightOnOCR-3-0.8B-IQ2_XS.gguf | IQ2_XS | 0.34 GB | Experimental. |

| LightOnOCR-3-0.8B-IQ2_XXS.gguf | IQ2_XXS | 0.33 GB | Experimental. |

| LightOnOCR-3-0.8B-Q2_0.gguf | Q2_0 | 0.34 GB | Extreme, group-64. Included for completeness. |

| LightOnOCR-3-0.8B-IQ1_M.gguf | IQ1_M | 0.32 GB | Extreme. Expect substantial degradation. |

| LightOnOCR-3-0.8B-IQ1_S.gguf | IQ1_S | 0.31 GB | Extreme. Expect substantial degradation. |

| LightOnOCR-3-0.8B-Q1_0.gguf | Q1_0 | 0.27 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/LightOnOCR-3-0.8B-GGUF:Q4_K_M -p "Hello"

Or download one file and point at it directly:

huggingface-cli download NANI-Nithin/LightOnOCR-3-0.8B-GGUF LightOnOCR-3-0.8B-Q4_K_M.gguf --local-dir .
llama-cli -m LightOnOCR-3-0.8B-Q4_K_M.gguf -p "Hello"

---

Quantized with llama.cpp by AgentQuantix on 2026-10-12. Licensing follows the base model.