legifx/Fara1.5-GGUF-Q4_K_M

Fara1.5-9B Q4_K_M GGUF

About

GGUF-quantized version of Microsoft Fara1.5-9B, a multimodal computer-use agent (CUA) from Microsoft Research AI Frontiers.

Fara1.5-9B is a vision-language model specialized for browser automation. It observes web browsers through screenshots and performs actions (click, type, scroll, navigate) to complete tasks end-to-end. It is NOT a general-purpose chat model — it is an autonomous web agent.

Original Model

• Source: microsoft/Fara1.5-9B
• Architecture: Qwen3.5-9B (Hybrid Transformer + SSM, Qwen3_5ForConditionalGeneration)
• Parameters: 9B
• Context Length: 262,144 tokens
• License: MIT

Quantization Details

• Format: GGUF (llama.cpp compatible)
• Quantization: Q4_K_M (4-bit, mixed precision K-quant)
• Original size: ~18 GB (BF16 safetensors)
• Quantized size: ~5.3 GB
• BPW (bits per weight): ~5.02
• Conversion tool: llama.cpp b10089 (convert_hf_to_gguf.py + llama-quantize)

Hardware Requirements

• RAM/VRAM: Minimum ~6-7 GB for inference (5.3 GB model + overhead)
• GPU: NVIDIA with CUDA support recommended. CPU inference is possible but slow.
• Tested GPU: RTX 5070 Ti (16 GB VRAM)

Usage with llama.cpp

# Server mode
LD_LIBRARY_PATH=/path/to/llama-bins \
  ./llama-server \
  -m Faral-1.5-GGUF-Q4_K_M.gguf \
  --host 0.0.0.0 \
  --port 8080 \
  -ngl 99 \
  -c 8192

# CLI mode
./llama-cli -m Faral-1.5-GGUF-Q4_K_M.gguf -p "Hello" -n 256

Usage with LM Studio

1. Download the GGUF file from this repository
2. Open LM Studio
3. Go to "My Models" → drag and drop the file, or browse to it
4. Select the model and start chatting

Note: LM Studio's vision/multimodal support may be limited. For full functionality (screenshot input), use llama.cpp with the multimodal CLI (llama-mtmd-cli).

Usage with Ollama

# Create a Modelfile
echo 'FROM ./Faral-1.5-GGUF-Q4_K_M.gguf' > Modelfile
ollama create fara1.5-9b -f Modelfile
ollama run fara1.5-9b

System Prompt (Recommended)

For optimal results, use the official system prompt from Microsoft:

You are Fara, a computer use agent (CUA) specialized for web browsers.
You are developed by Microsoft AI Frontiers. You assist users with
completing and automating tasks that require the use of a web browser.

Key Features

• Vision-only perception: Operates on browser screenshots alone, no DOM access needed
• Coordinate-grounded actions: Predicts pixel-level click targets directly
• Safety design: Trained to pause before irreversible actions (payments, form submissions, account logins)
• 262K context: Long enough for multi-step browser task trajectories
• Tool-call output: Emits structured XML <tool_call> blocks with actions like left_click(x,y), type(text), scroll(), visit_url(), etc.

Supported Actions

Action	Purpose
left_click, right_click, double_click	Mouse clicks at (x, y) coordinates
type, key	Keyboard input
scroll, hscroll	Page scrolling
visit_url, history_back	Browser navigation
web_search	Search query
terminate	End task with final answer

Limitations

• This is a specialized web agent, not a general chatbot
• Vision/multimodal input required for full functionality (screenshots)
• Best used with 1440×900 screen resolution
• Without vision capabilities (text-only mode), quality degrades significantly
• Knowledge cutoff: April 2026

Credits

• Model: Microsoft Research AI Frontiers (microsoft/Fara1.5-9B)
• Quantization: llama.cpp project (ggerganov/llama.cpp)
• GGUF conversion by: legifx

License

MIT License — same as the original model.