Fara 7B — Quantized by BatiAI

Microsoft’s agentic multimodal model in a 7B package. imatrix-calibrated GGUF quantizations of the official microsoft/Fara-7B (Qwen 2.5 VL backbone, MIT), released 2026-05-19 by Microsoft Research. Free, unlimited, on-device AI for Mac via BatiFlow.

Multimodal-capable (vision) via separate mmproj on Hugging Face. Ollama ships text-only.

Available tags

Tag	Size	Min RAM	Use case
`:iq3`	3.1 GB	8 GB	Smallest footprint — Mac mini M4 16GB OK
`:q3`	3.8 GB	8 GB	K-quant alt for iq3
`:iq4`	4.2 GB	10 GB	Best size/quality ratio
`:q4`	4.7 GB	10 GB	Recommended for most users
`:q5`	5.4 GB	12 GB	Higher fidelity
`:q6`	6.3 GB	14 GB	Near-original quality
`:q8`	8.1 GB	16 GB	Original-grade

All 7 are imatrix-calibrated (wikitext-2-raw).

Quick Start

ollama pull batiai/fara-7b:q4
ollama run  batiai/fara-7b:q4

Why Fara 7B?

Microsoft Research’s agentic multimodal model built on Qwen 2.5 VL backbone. Designed for screen understanding + agent flows (web automation, UI navigation, document analysis).

7B parameters — runs comfortably on Mac mini M4 16GB
Qwen 2.5 VL backbone — strong vision + multilingual base
128K context — long documents, multi-turn screen sessions
MIT license — fully commercial-friendly
arxiv:2511.19663 — research paper available

RAM guide

Your Mac	`:iq3` 3GB	`:q3` 4GB	`:iq4` 4GB	`:q4` 5GB	`:q5` 5GB	`:q6` 6GB	`:q8` 8GB
16 GB	✅	✅	✅	✅	✅	⚠ tight	❌
24 GB	✅	✅	✅	✅	✅	✅	✅
32 GB+	✅	✅	✅	✅	✅	✅	✅

16 GB Mac sweet spot: :q4 or :q5. Smallest footprint at this quality tier in the BatiAI catalog.

Vision usage (multimodal)

Ollama ships text-only. For image input, use llama.cpp with the separate mmproj:

hf download batiai/Fara-7B-GGUF --include "*Q4_K_M*" --include "mmproj-*-Q6_K.gguf" \
    --local-dir ./fara-7b

llama-mtmd-cli \
    -m ./fara-7b/microsoft-Fara-7B-Q4_K_M.gguf \
    --mmproj ./fara-7b/mmproj-microsoft-Fara-7B-Q6_K.gguf \
    --image input.jpg \
    -p "Describe this image."

Why BatiAI?

Quantized directly from official Microsoft BF16 weights — no re-quantization chains
7-quant variants (IQ3/Q3/IQ4/Q4/Q5/Q6/Q8) for every Mac tier
BatiAI metadata signed (general.author=BatiAI, general.url=https://flow.bati.ai)
Multimodal mmproj packaged alongside on Hugging Face

License

Inherits source: MIT.

Built for BatiFlow

flow.bati.ai — free, on-device AI automation for Mac. 5 MB app, 100 % local, unlimited.

Details

Readme