---

license: apache-2.0

base_model:

• bravesoftware/Ocelot-1-VL
• Qwen/Qwen3-VL-4B-Instruct

pipeline_tag: image-text-to-text

tags: [gguf, llama.cpp, qwen3-vl, vision, summarization]

---

Ocelot-1-VL GGUF

Merged and quantized GGUF builds of bravesoftware/Ocelot-1-VL, a Qwen3-VL-4B LoRA specialized for webpage summarization from page text or screenshots.

The Ocelot LoRA was merged into Qwen/Qwen3-VL-4B-Instruct in BF16 before conversion. The vision projector is separate and is unchanged by the language-side LoRA.

Intended use

This is not a general-purpose assistant. Use it only to summarize webpage text or webpage screenshots with the training-aligned prompt. See the original model card for the full prompt contract, limitations, and safety guidance.

The is the text of a webpage: <page>
... page plain text here ...
</page>
Summarise the content between the <page> tags, or if no content is found use the screenshots provided, in the Brave Summary style.

For screenshots, begin with The following is a screenshot of a webpage: and append the same fixed instruction.

Files

| Quant | Approx. size | Notes |

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

| BF16 | 7.5 GB | Maximum fidelity |

| Q8_0 | 4.0 GB | Near-BF16 quality |

| Q6_K | 3.1 GB | High quality |

| Q5_K_M / Q5_K_S | 2.7 / 2.6 GB | 5-bit variants |

| Q4_K_M / Q4_K_S | 2.3 / 2.2 GB | Recommended balance / smaller |

| Q4_1 / Q4_0 | 2.4 / 2.2 GB | Legacy compatibility |

| IQ4_NL / IQ4_XS | 2.2 / 2.1 GB | Non-linear 4-bit variants |

| Q3_K_L / Q3_K_M / Q3_K_S | 2.1 / 1.9 / 1.8 GB | 3-bit variants |

| Q2_K | 1.6 GB | Maximum compression, largest quality loss |

| mmproj-Q8_0 / mmproj-BF16 | 433 / 813 MB | Vision projectors |

Run

llama-cli \
  -m Ocelot-1-VL-Q4_K_M.gguf \
  -p 'The is the text of a webpage: <page>Page text here</page> Summarise the content between the <page> tags, or if no content is found use the screenshots provided, in the Brave Summary style.' \
  -n 512 -c 8192

For screenshots, add --mmproj mmproj-Ocelot-1-VL-Q8_0.gguf --image webpage.png and use the screenshot prompt described above.

Conversion and validation

• Adapter revision: fbc473443eec370c3b551363556ab75d2b485db6
• Base revision: ebb281ec70b05090aa6165b016eac8ec08e71b17
• llama.cpp revision: 0278d8362d78c5de291bc03b76016f7f74b2ab77
• All 15 language files passed generation smoke tests.
• Q4_K_M with the Q8 projector passed screenshot summarization.
• Apple M4 Pro Q4 benchmark: 313 prompt tokens/s and 26 generation tokens/s with Metal.

Quantization can reduce output quality, especially at 2 and 3 bits. Validate against your own webpage corpus before production use.