Holo3.1 ships quantized agents that run locally on your laptop

Hcompany just dropped Holo3.1, and it’s not a minor point release. This is the first time the company has shipped quantized checkpoints—FP8, Q4 GGUF, and NVFP4—designed specifically for local inference. Translation: you can now run a capable computer-use agent entirely on your own hardware, with nothing leaving your network.

The performance numbers are the real story here. On AndroidWorld, the 35B-A3B model jumps from 67% to 79.3%. Even the smaller 4B and 9B variants climb from 58% to 72%. That’s not incremental. The team also claims a 25% improvement over Holo3 when evaluated inside their own Holotab product harness. And for the first time, function-calling protocols are supported natively, achieving near-parity with the structured JSON outputs that were already available. If you’re building inside a third-party agent stack, that’s a meaningful friction point removed.

Speed is where things get interesting. On a DGX Spark, NVFP4 in W4A16 configuration delivers 1.41× the token throughput of FP8 and 1.74× that of full-precision BF16. Agent harness optimizations done with NVIDIA compound those gains, cutting average step time from 6.8 seconds to 3.3 seconds—roughly a 2× end-to-end speedup. The Q4 GGUF checkpoints target consumer hardware directly, with reference numbers for Apple Silicon included.

I’m watching two things here. First, how much model degradation actually occurs—the team says FP8 and NVFP4 scores are “only about two points below” BF16 on OSWorld, which is a tradeoff most production deployments will happily take. Second, whether the mobile gains hold up outside AndroidWorld’s benchmark suite. The 79.3% number is impressive, but mobile automation is notoriously brittle in the wild. Still, shipping four model sizes (0.8B, 4B, 9B, and 35B-A3B) alongside quantized weights signals Hcompany is serious about deployment flexibility, not just benchmark bragging.