File size: 2,805 Bytes
37ebf62 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
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](https://huggingface.co/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](https://huggingface.co/bravesoftware/Ocelot-1-VL) for the full prompt contract, limitations, and safety guidance.
```text
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
```bash
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.
|