---

base_model: microsoft/Fara1.5-27B

library_name: gguf

license: mit

pipeline_tag: image-text-to-text

language:

• en

tags:

• gguf
• quantized
• llama.cpp
• computer-use
• cua
• web-agent
• multimodal
• vision-language
• agent

---

Fara1.5-27B - GGUF Quants & Multimodal Projectors

This repository contains GGUF quantizations and multimodal projector (mmproj) files for microsoft/Fara1.5-27B.

• Original Model: microsoft/Fara1.5-27B
• Developer: Microsoft Research AI Frontiers
• Base Architecture: Qwen3.5-27B (Multimodal Vision-Language / Computer Use Agent)
• Quantization Format: GGUF (Q3_K_M, Q4_K_M, Q4_K_S, Q5_K_M, Q5_K_S, Q6_K, Q8_0) + mmproj vision projectors

---

Repository Files

Model Weights (.gguf)

| File Name | Size | Quant Method | Description |

|---|---|---|---|

| Fara1.5-27B.Q3_K_M.gguf | ~13.3 GB | Q3_K_M | 3-bit medium. Lowest VRAM footprint. |

| Fara1.5-27B.Q4_K_S.gguf | ~15.6 GB | Q4_K_S | 4-bit small. Balanced size and performance. |

| Fara1.5-27B.Q4_K_M.gguf | ~16.5 GB | Q4_K_M | 4-bit medium. Recommended balance of speed, memory usage, and quality. |

| Fara1.5-27B.Q5_K_S.gguf | ~18.7 GB | Q5_K_S | 5-bit small. High quality, lower loss. |

| Fara1.5-27B.Q5_K_M.gguf | ~19.2 GB | Q5_K_M | 5-bit medium. Superior precision. |

| Fara1.5-27B.Q6_K.gguf | ~22.1 GB | Q6_K | 6-bit quantization. Extremely close to native FP16 accuracy. |

| Fara1.5-27B.Q8_0.gguf | ~28.6 GB | Q8_0 | 8-bit quantization. Maximum accuracy. |

Multimodal Vision Projectors (.mmproj-*.gguf)

(Required to pass screenshot images into the model in llama.cpp)

| File Name | Size | Format | Description |

|---|---|---|---|

| Fara1.5-27B.mmproj-q8_0.gguf | ~629 MB | Q8_0 | Quantized vision projector (recommended for low VRAM). |

| Fara1.5-27B.mmproj-bf16.gguf | ~931 MB | BF16 | Native BFloat16 vision projector. |

| Fara1.5-27B.mmproj-f16.gguf | ~931 MB | FP16 | Standard FP16 vision projector. |

---

Quickstart Guide

Because Fara1.5-27B processes screenshots, you must pass both a model weight file (-m) AND a multimodal projector file (--mmproj) during inference in llama.cpp.

1. Running with llama.cpp

# Clone and build llama.cpp
git clone [https://github.com/ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp)
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build --config Release -j

# Download the model and vision projector from this repository
huggingface-cli download Abiray/Fara1.5-27B-GGUF Fara1.5-27B.Q4_K_M.gguf --local-dir .
huggingface-cli download Abiray/Fara1.5-27B-GGUF Fara1.5-27B.mmproj-f16.gguf --local-dir .

# Serve with llama-server
./build/bin/llama-server \
  -m Fara1.5-27B.Q4_K_M.gguf \
  --mmproj Fara1.5-27B.mmproj-f16.gguf \
  -ngl 99 \
  --port 8000 \
  -c 262144