Hcompany's Holo3.1 delivers 79.3% mobile accuracy with local AI agents
New quantized models run 1.74x faster on consumer hardware, matching near-full precision scores.
Deep Dive
Hcompany releases Holo3.1, a major update to its computer-use agent family. The new models achieve 79.3% on AndroidWorld (up from 67% for the 35B-A3B variant) and ship quantized checkpoints (FP8, Q4 GGUF, NVFP4) for local inference. On DGX Spark, NVFP4 delivers 1.74× the total token throughput of BF16. New smaller 0.8B, 4B, and 9B sizes enable cost-effective and private deployment.
Key Points
- Mobile automation on AndroidWorld rises to 79.3% for 35B-A3B (up from 67%) and 72% for smaller models (up from 58%).
- First quantized checkpoints: FP8, Q4 GGUF, NVFP4; NVFP4 delivers 1.74x speedup over BF16 on DGX Spark.
- New model sizes (0.8B, 4B, 9B) plus native function-calling support enable local/private deployment across agent frameworks.
Why It Matters
Enables private, fast computer-use agents on consumer hardware without sacrificing accuracy, expanding automation to mobile and local workflows.