Fara1.5 27B
microsoft/Fara1.5-27B
Open Source · chat · open-weights
Context
262.1K
Max output
—
Pricing
Open weights
Modalities
text · image
Released
17 Jul 2026
License: mit · microsoft/Fara1.5-27B
AI summary
● machine-written
Fara1.5-27B: 27B parameter computer use agent on Qwen3.5
Fara1.5-27B is a 27-billion parameter computer use agent built on Qwen3.5, designed to complete complex browser-based tasks through an observe-think-act loop. The model achieves 72.3% on Online-Mind2Web benchmark, outperforming larger proprietary systems. It is available as open-weights under MIT license and can run on single A100/H100/B200 GPUs.
What's new
- Fara1.5 family released at three scales (4B, 9B, 27B) trained on FaraGen1.5 pipeline
- 262K token context window for multi-screenshot trajectories
- 72.3% task success rate on Online-Mind2Web, outperforming larger proprietary systems
- Trained on trajectory data including synthetic domains for tasks beyond live websites
- Open-weights release under MIT license on Hugging Face
Best for
Browser automation and web agent tasksForm filling and cross-site comparison shoppingComplex multi-step web interactionsDeployment on modest single-GPU hardware