Holo3.1: A New Generation of Fast, Local Computer-Use Agents

H has released Holo3.1, a new family of computer-use agents designed to operate reliably across desktop, mobile, and web environments while supporting multiple agent frameworks and deployment targets—from cloud inference to fully local execution.

The update addresses a key production challenge: models that perform well in controlled settings often struggle when deployed across different platforms and harnesses. Holo3.1 tackles this through improvements in robustness across three dimensions: diverse environments, agent framework compatibility, and flexible deployment options.

Mobile and Cross-Platform Gains#

Building on the Qwen family, Holo3.1 significantly expands mobile automation capabilities. On AndroidWorld benchmarks, the 35B-A3B model jumps from 67% to 79.3% performance, while smaller 4B and 9B variants improve from 58% to 72%.

The model now supports function-calling protocols alongside structured JSON outputs, enabling near-parity performance across different agent stacks and execution frameworks. Within H's Holotab product harness, Holo3.1 delivers over 25% improvement compared to its predecessor.

Quantized Models for Local Deployment#

For the first time, H is releasing quantized checkpoints optimized for edge and local inference:

• FP8 and NVFP4: Using NVIDIA's Model Optimizer in W4A16 configuration, these achieve similar OSWorld scores to full-precision BF16 models (only ~2 points lower) while delivering 1.41× and 1.74× throughput gains respectively.
• Q4 GGUF: Designed for consumer hardware (Windows, Mac, Apple Silicon), enabling fully private, local agent execution with nothing leaving the user's network.

On DGX Spark, agent harness optimizations combined with NVFP4 quantization deliver approximately 2× end-to-end speedup, cutting average step time from 6.8s to 3.3s.

Model Sizes and Availability#

Holo3.1 comes in four sizes—0.8B, 4B, 9B, and 35B-A3B—balancing performance against cost and deployment constraints. All quantized variants are available via the Holo Models API and Hugging Face.

The release represents progress toward H's goal of universal computer-use agents that operate seamlessly across any environment, integrate with any agent framework, and run wherever users need them.