Downloads · 30 days
111
11% of all-time downloads
runanywhere/Fara1.5-4B-GGUF
Fara1.5-4B-GGUF is a image-text-to-text model from runanywhere. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for gguf. The card lists the license as mit.
llama.cpp GGUF conversion of microsoft/Fara1.5-4B, Microsoft AI Frontiers' agentic Computer-Use Agent (CUA) for web browsers (Qwen3.5-VL arch, Qwen35ForConditionalGeneration). This repo only re-packages Microsoft's mo…
Downloads · 30 days
111
11% of all-time downloads
All-time downloads
998
Public
Repo size
11.8 GB
Likes
0
Public
Click a slice to open those files.
.gguf11.8 GB · 100%
From the Hugging Face model README
llama.cpp GGUF conversion of
microsoft/Fara1.5-4B, Microsoft AI
Frontiers' agentic Computer-Use Agent (CUA) for web browsers (Qwen3.5-VL arch,
Qwen3_5ForConditionalGeneration). This repo only re-packages Microsoft's model for
on-device inference — all capabilities and credit belong to Microsoft AI Frontiers.
Original model, full system prompt, and MagenticLite agent scaffold:
https://huggingface.co/microsoft/Fara1.5-4B and https://github.com/microsoft/fara.
| File | What | Size |
|---|---|---|
Fara1.5-4B-Q4_K_M.gguf | text model, 4-bit (Q4_K_M) | ~2.5 GB |
Fara1.5-4B-f16.gguf | text model, full precision (f16) | ~7.9 GB |
mmproj-Fara1.5-4B-f16.gguf | vision projector (required for images) | ~0.6 GB |
A VLM needs both a text GGUF and the mmproj. llama.cpp maps the vision tower to
PROJECTOR_TYPE_QWEN3VL and the text model to LLM_ARCH_QWEN3VL.
Produced with a qwen3_5-aware llama.cpp convert_hf_to_gguf.py:
python convert_hf_to_gguf.py microsoft/Fara1.5-4B --outfile Fara1.5-4B-f16.gguf --outtype f16
python convert_hf_to_gguf.py microsoft/Fara1.5-4B --mmproj --outfile mmproj-Fara1.5-4B-f16.gguf
llama-quantize Fara1.5-4B-f16.gguf Fara1.5-4B-Q4_K_M.gguf Q4_K_M
llama-mtmd-cli -m Fara1.5-4B-Q4_K_M.gguf --mmproj mmproj-Fara1.5-4B-f16.gguf \
--image screenshot.png -p "<FARA_SYSTEM_PROMPT> ... your task"
Fara is a web-browser agent. Use Microsoft's contract (see the base model):
computer_use tool schema inside
<tools></tools>.<tool_call>{"name": "computer_use", "arguments": {...}}</tool_call>.x*1.44, y*0.9 for a 1440×900 window). 1440×900 is the recommended capture size.Verified running on Apple silicon through llama.cpp (CPU and Metal GPU). On real 1440×900
browser screenshots with Fara's exact CUA prompt, the Q4_K_M produced valid left_click
tool-calls with correct coordinate grounding (matching an independent community Q4_K_M
byte-for-byte on shared tasks, and matching the MLX 4-bit build).
MIT, inherited from microsoft/Fara1.5-4B. Review Microsoft's model card for intended use, limitations, and responsible-AI guidance (Fara pauses at safety-critical points; keep a human in the loop for irreversible actions).