Holo3.1 Brings Computer-Use Agents to Local Devices and Mobile

Hugging Face has released Holo3.1, an updated computer-use agent family designed to run on local devices and mobile platforms alongside cloud deployment. According to the Hugging Face Blog, the release includes quantized checkpoints optimized for on-device inference—including FP8, Q4 GGUF, and NVFP4 formats—and addresses a fundamental challenge teams encountered deploying the original Holo3: strong performance in controlled evaluation settings often failed to transfer to real-world production environments across different hardware, frameworks, and operating systems.

Mobile Automation Gains

The most significant improvement targets mobile execution. According to Hugging Face, the 35B-A3B model achieves 79.3% on AndroidWorld, up from 67% with Holo3—a 12.3-point gain. Smaller variants (4B and 9B parameters) improved from 58% to 72%, demonstrating that mobile automation gains hold across the model family. This expansion moves computer-use agents beyond browser and desktop workflows, where they historically concentrated.

Cross-Framework Compatibility

Holo3.1 addresses a secondary pain point: agent framework heterogeneity. Hugging Face reports that the new model introduces native function-calling protocol support alongside the structured JSON outputs available in Holo3. Across OSWorld and internal benchmarks covering e-commerce, business software, and collaboration tools, function-calling and native execution now achieve near-parity performance. Within the Holotab product harness specifically, Holo3.1 delivers more than a 25% improvement over Holo3. This compatibility layer reduces friction for teams integrating computer-use capabilities into third-party agent stacks that may not natively support Hugging Face’s output format.

Quantization and Model Sizing

To enable cost-effective local deployment, Hugging Face is releasing new model sizes including 0.8B, 4B, and 9B parameter variants alongside the flagship 35B-A3B model. The inclusion of quantized checkpoints—particularly Q4 GGUF, which targets CPU-based and consumer GPU inference—lowers the computational barrier for on-device execution. This is a departure from Holo3, which did not emphasize local-inference optimization.

Why This Matters

Holo3.1’s release reflects a maturation in computer-use agent adoption. As teams move from prototyping to production at scale, the ability to run agents locally addresses both privacy concerns (workflows never leave the device) and operational costs (inference runs without cloud API calls). Teams deploying agents across heterogeneous environments—web dashboards, desktop applications, mobile apps—now have a single model family designed to handle distribution shift across all three. Organizations evaluating computer-use vendors should factor in deployment flexibility and framework compatibility alongside raw benchmark scores; Holo3.1’s quantized variants and cross-harness support reduce the hidden integration cost that often outweighs raw performance gains in production decisions.