Fara-7B

Fara-7B: An Efficient Agentic Model for Computer Use Official Microsoft Blog Technical Report Paper Github Try Fara-7B on Microsoft Foundry Model Summary Developer:** Microsoft Research Description:** Fara-7B is Microsoft's first agentic small language model (SLM) designed specifically for computer use. With only 7 billion parameters, Fara-7B is an ultra-compact Computer Use Agent (CUA) that achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems. Model Architecture:** Multimodal decoder-only language model that takes an image (screenshot) + text context. It directly predicts thoughts and actions with grounded arguments. Current production baselines leverage Qwen 2.5-VL (7B). Parameters:** 7 Billion Inputs:** User goal (text), current screenshot(s), history of previous outputs (thoughts + actions text) from the agent. Context Length:** 128k Outputs:** Generated text in response to the input, with a chain-of-thought block followed by a tool call block to indicate the action. GPUs:** 64 H100s Training Time:** 2.5 days Public Data Summary:** N/A Dates:** Trained between 26th October 2025 to 29th October 2025 Status:** Static model trained on public and private data Release Date:** November 24th, 2025 License:** MIT Model Dependencies:** Qwen 2.5 VL Additional Assets:** N/A Acceptable Use Policy:** N/A 1. Model Overview Fara is a 7B Computer Use Agent (CUA) model specialized for taking actions on the web to accomplish high-level user tasks. Beyond understanding webpage layout and basic action mechanics, it plans and executes high-level goals like booking restaurants, applying for jobs, planning trips, and buying shopping lists. Its training relies on a large-scale, fully synthetic dataset of action trajectories generated and verified by a multi-agent pipeline. Fara perceives browser inputs via screenshots, while internal reasoning and state history are recorded textually. Based on recent screens