FlowtandBriefAI · Software · ResearchMeet Flowtand

Holo3.1 adds mobile gains, function calling and quantized local checkpoints for computer-use agents

Announcement

Hugging Face

Browse Agents topic page →

Hcompany released the Holo3.1 family of computer-use models, adding mobile environment support, native function-calling protocols and, for the first time in the line, quantized checkpoints (FP8, Q4 GGUF, NVFP4) aimed at local inference. The publisher reports AndroidWorld gains, near-parity between function calling and native execution, and speedups on DGX Spark, with sizes from 0.8B to 35B-A3B.

What Hcompany says Holo3.1 changes

Hcompany describes Holo3.1 as a refresh of its Holo3 computer-use model family, built on the Qwen family and positioned around robustness rather than raw benchmark gains alone. Three areas are named: GUI environments, agent harnesses and deployment targets. On environments, the publisher reports that its 35B-A3B model improves AndroidWorld performance from 67% to 79.3%, while the 4B and 9B variants move from 58% to 72%, extending coverage toward mobile. On harnesses, Holo3.1 adds native function-calling protocols alongside the structured JSON outputs Holo3 already offered; Hcompany says function calling and native execution reach near-parity on OSWorld and its internal suite, and reports a >25% improvement over Holo3 inside its Holotab harness. On deployment, the release ships quantized 35B-A3B checkpoints in FP8, Q4 GGUF and NVFP4. Hcompany states FP8 and NVFP4 match each other on OSWorld, about two points below BF16, and reports on DGX Spark that NVFP4 reaches 1.41x the total token throughput of FP8 and 1.74x that of BF16. A compound ~2x end-to-end speedup is also reported, with average step time falling from 6.8s to 3.3s.

A scenario for a desktop automation engineer

Consider an engineer maintaining an internal agent that drives a mix of browser SaaS tools and native desktop applications for back-office staff. A plausible use of the evidence is to test whether a smaller Holo3.1 variant can replace a remote inference call for simpler workflows, while the 35B-A3B quantized checkpoint handles harder multi-step tasks on a local machine or a nearby DGX Spark. Because Hcompany says the agent can run locally while the model runs on the same machine or on a Spark on the same network, teams with data-residency constraints could try keeping both inside the network boundary. The function-calling addition matters here too: an engineer whose harness already expects tool-call schemas could try integrating Holo3.1 through that path rather than rewriting around JSON outputs, then compare behavior on the team's own scripts. These are hypotheses to test against the publisher's claims, not established outcomes.

How a team could trial it

A team could start by recording a baseline on its own automation scripts with the currently deployed model, measuring task success rate, average step time and failure categories across web, desktop and, if relevant, mobile targets. Then it could run the same scripts with Holo3.1 variants, including at least one quantized 35B-A3B checkpoint and one smaller model, and compare on identical hardware and network conditions. Success criteria should be set in advance: for example, no drop in task success beyond an agreed threshold, and a step-time improvement large enough to justify switching. Failure checks should include watching for new error patterns on mobile or unfamiliar harnesses, and re-running a sample of past failures to see whether they persist. Because these are GUI agents operating on real interfaces, a human should review any high-impact actions such as form submissions or purchases during the trial, and the team should log model decisions for later audit.

Trade-offs, prerequisites and when to hold off

The main trade-off is precision versus footprint: Hcompany reports FP8 and NVFP4 scoring about two points below BF16 on OSWorld, so teams with little tolerance for accuracy loss may prefer full precision where hardware allows. Local deployment requires suitable hardware and, for the Q4 GGUF path, a consumer-grade Windows or Mac machine or a DGX Spark reachable on the same network; the excerpt gives reference numbers only for Apple Silicon and Spark, so other configurations are unproven from this source. The published quantized checkpoints are described for 35B-A3B, and the pricing and licensing terms of the models API are not covered here. Teams should not assume that the reported mobile and cross-harness gains transfer to their own environments. If a workflow depends on strict reproducibility, regulated actions without human review, or hardware outside the reported setups, it may be wiser to wait for independent evaluation before adopting.

Context and limitations

  • Based on a short source excerpt; the full article was not analyzed.
  • All benchmark and speed figures are Hcompany's own reported numbers, not independently verified results.
  • Quantized checkpoints are described only for 35B-A3B; per-size performance and hardware coverage are not detailed in the excerpt.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.