Holo3.1: Fast & Local Computer Use Agents

After releasing Holo3, the team found that strong benchmark performance alone wasn’t enough for production. Users needed robustness across diverse environments—web, desktop, mobile—and compatibility with different agent harnesses. The core tension was that performance in one setting didn’t transfer to another, and teams faced distribution shifts when moving from evaluation to real-world deployment, especially as demand grew for fully local execution on consumer hardware.

Holo3.1 addresses these gaps across three dimensions. On mobile, the 35B-A3B model jumps from 67% to 79.3% on AndroidWorld. It adds native function-calling protocols alongside JSON outputs, achieving near-parity across harnesses and a 25% improvement inside Holotab. New small sizes (0.8B, 4B, 9B) enable cost-performance tradeoffs. Quantized checkpoints—FP8, Q4 GGUF, and NVFP4—are released for the first time. On DGX Spark, NVFP4 delivers 1.41× the token throughput of FP8 and 1.74× that of BF16, and the combined optimizations cut average step time from 6.8s to 3.3s, a ~2× end-to-end speedup.

For builders, Holo3.1 signals a shift from chasing benchmark scores to engineering for deployment flexibility. The quantized checkpoints and smaller models make private, low-latency computer-use agents practical on consumer devices. By improving cross-environment robustness and framework integration, this release lowers the barrier to building agents that work across desktop, mobile, and multiple agent stacks—a serious step toward universal computer-use agents.