Holo3.1: AndroidWorld Jumps From 67% to 79.3%, First Computer-Use Model With Local Quantized Weights
Computer-use agents have been GPU-cluster-only since the category emerged. Holo3.1 is the first release to change that.
HCompany released Holo3.1 on June 2, expanding the original Holo3 lineup to four sizes and shipping quantized weights — FP8, Q4 GGUF, and NVFP4 — for the first time. The result is a computer-use model family that runs locally on consumer hardware without meaningful accuracy loss.
Benchmark Gains
The flagship 35B-A3B model improves AndroidWorld from 67% to 79.3% — a 12-point jump. Smaller models (4B, 9B) move from 58% to 72%.
AndroidWorld measures an agent’s ability to complete real Android tasks: navigating apps, filling forms, responding to notifications, composing messages. A 12-point gain at the top of the size range is operationally significant, not a benchmark artifact.
Across OSWorld, HCompany’s internal H Corporate benchmarks, and Holotab — the company’s product harness — Holo3.1 delivers a consistent 25%+ improvement over Holo3. Function-calling and native execution now reach near-parity performance, removing a key decision point for developers choosing between interface paradigms.
Local Inference: Quantized Checkpoints for Consumer Hardware
Holo3.1 ships three quantization formats for the 35B-A3B model:
- FP8: Standard 8-bit float, broadly compatible with modern inference stacks
- NVFP4 (W4A16): 4-bit weights, 16-bit activations, produced with NVIDIA’s Model Optimizer — delivers 1.41x throughput over FP8 and 1.74x over BF16
- Q4 GGUF: CPU-compatible 4-bit quantization for local inference on consumer GPUs via llama.cpp and compatible runtimes
On NVIDIA DGX Spark, combining NVFP4 quantization with agent harness optimizations from a joint HCompany-NVIDIA project produces a compound 2x end-to-end speedup — average step time drops from 6.8s to 3.3s. That is the difference between a computer-use agent that runs a 20-step workflow in under two minutes versus over three.
OSWorld scores under FP8 and NVFP4 land roughly 2 points below the full-precision BF16 checkpoint — acceptable accuracy degradation for a 1.74x throughput gain.
Model Family
| Size | Target use case |
|---|---|
| 0.8B | Ultra-lightweight local agents |
| 4B | Cost-efficient deployment |
| 9B | Balanced performance and latency |
| 35B-A3B | State-of-the-art performance |
The 4B and 9B sizes are new with this release. Holo3 shipped as a single flagship. The expanded ladder makes it practical to deploy computer-use capability across a wider range of hardware and cost constraints.
Why Local Inference Matters for Computer Use
Computer-use agents run in tight feedback loops: capture screen, send to model, receive action, execute, repeat. Every round-trip to a cloud API adds latency that compounds across hundreds of steps in a long session. A local quantized model on DGX Spark or a high-end workstation can run that loop at 3.3 seconds per step — competitive with cloud round-trips at peak demand, without the per-token cost.
The Q4 GGUF checkpoints push the question further: whether a well-quantized 35B can run adequately on a consumer RTX 4090 is a question Holo3.1 now makes answerable.
Holo3.1 is available via API now. Model weights for Hugging Face are included in the release. An upcoming desktop agent harness is in development.