---

license: mit

library_name: gguf

pipeline_tag: image-text-to-text

language:

• en

base_model: microsoft/Fara1.5-9B

tags:

• gguf
• qwen3_5
• computer-use
• cua
• web-agent
• multimodal
• vision-language
• browser-automation
• fara
• llama.cpp

---

Fara1.5-9B GGUF

GGUF conversions of microsoft/Fara1.5-9B, Microsoft's 9B multimodal computer-use agent for screenshot-driven browser automation.

The files were converted from source revision 8238b75529e98434621204024b149860b8071d33 and validated with llama.cpp build b10066 (86a9c79f8).

Files

| File | Size | Description |

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

| Fara1.5-9B-Q4_K_M.gguf | 5.24 GiB | Recommended balance of size and quality; 5.02 BPW |

| Fara1.5-9B-Q8_0.gguf | 8.87 GiB | Higher-fidelity 8-bit quantization; 8.50 BPW |

| Fara1.5-9B-mmproj-F16.gguf | 0.855 GiB | F16 vision projector required for screenshot/image input with either quant |

The projector is shared by both text-model quantizations. Download it alongside the selected model file for multimodal use.

llama.cpp usage

Fara is a screenshot-driven computer-use model. For image grounding, use at least 1,024 image tokens:

llama-mtmd-cli \
  -m Fara1.5-9B-Q4_K_M.gguf \
  --mmproj Fara1.5-9B-mmproj-F16.gguf \
  --image screenshot.png \
  --image-min-tokens 1024 \
  -ngl 99 \
  -p "Describe the next browser action."

Server example:

llama-server \
  -m Fara1.5-9B-Q4_K_M.gguf \
  --mmproj Fara1.5-9B-mmproj-F16.gguf \
  --image-min-tokens 1024 \
  -ngl 99 \
  --ctx-size 8192

For full computer-use behavior, follow the original model card's system prompt, tool schema, screenshot-history format, and critical-point safety requirements. Microsoft recommends running the browser in a sandbox with domain allow-lists, monitoring, and an immediate pause control. The model was most commonly trained with 1440 x 900 screenshots.

Conversion and validation

• Source: four BF16 safetensor shards, 760 indexed tensors, with no missing, extra, or misassigned tensors.
• Text GGUF architecture: qwen35, 427 tensors.
• Vision projector architecture: clip, 334 tensors.
• Quantization: standard llama-quantize Q8_0 and Q4_K_M without an importance matrix.
• Runtime: both quantizations loaded and generated successfully on an RTX 4090.
• Multimodal: Q4_K_M plus the F16 projector successfully encoded and interpreted a repository screenshot.

SHA-256

83198b9f854655de273841116135bd7d12ea763c0fab6c198fb91de3b5fe4441  Fara1.5-9B-Q8_0.gguf
8cd4370324f5477ebd48b59ad32bdc1edc19dcb8dcf21479c7bebd3ed572dc7f  Fara1.5-9B-Q4_K_M.gguf
6b9299f56b6c1cc0eab1e3404664137bdde4dbd139da2ead59502d7da340547c  Fara1.5-9B-mmproj-F16.gguf

Fara1.5-9B is released under the MIT license. See the original model card for intended use, limitations, safety guidance, evaluation results, and attribution.