TL;DR: H Company released Holo3.1 on June 2, 2026 — the first production-grade computer-use model family to ship quantized weights (FP8, NVFP4, Q4 GGUF) for fully local deployment. The 35B-A3B model now scores 79.3% on AndroidWorld (up from 67%), and agents can run entirely on-device with as little as 12GB of VRAM.
79.3%AndroidWorld (35B-A3B)
2×Speed gain with NVFP4 vs FP8
140msPerception-to-action on RTX 4090
12GBMin VRAM (4-bit quant)

What Is Holo3.1?

H Company's Holo3.1 is an updated family of Vision-Language Models (VLMs) designed specifically for computer-use agents — systems that perceive a screen and autonomously operate a keyboard and mouse to complete real-world tasks. The original Holo3 (released March 2026) set a new state-of-the-art on OSWorld-Verified at 78.85%, outperforming GPT-5.4 and Opus 4.6 at a fraction of the cost. Holo3.1 builds on that foundation with three major production improvements.

Three Key Improvements

1. Local Quantized Inference

For the first time, H Company ships quantized checkpoints alongside the full-precision model. The 35B-A3B model is available in FP8, NVFP4, and Q4 GGUF formats. On an NVIDIA DGX Spark, NVFP4 (W4A16) cuts average agent step time from 6.8s (FP8) to 3.3s — roughly a 2× end-to-end speedup. On an RTX 4090, perception-to-action latency reaches 140ms, four times faster than typical cloud API roundtrips. With 4-bit quantization, the full agent stack fits on a 12GB VRAM GPU.

The privacy implication is significant: continuous desktop screenshots never leave the machine. This opens computer-use agents to finance, healthcare, and legal workflows where cloud API transmission is prohibited.

2. Mobile Automation

Holo3.1 adds first-class mobile support. Trained on Android UI traces, the 35B-A3B model jumps from 67% to 79.3% on the AndroidWorld benchmark. The 4B and 9B variants go from 58% to 72%. In H Company's own Holotab product harness, Holo3.1 shows more than 25% improvement over Holo3.

3. Native Function Calling

Holo3.1 supports the OpenAI-compatible function-calling protocol alongside its existing structured JSON output. The two execution modes now perform at near-parity, eliminating the 10–15% accuracy gap that hampered Holo3 integration with third-party frameworks. This means Holo3.1 plugs directly into LangGraph, CrewAI, AutoGen, and custom harnesses without adapter layers.

Model Family

Model Parameters Best For
Holo3.1-0.8B 0.8B Ultra-lightweight edge agents
Holo3.1-4B 4B Cost-efficient private deployment
Holo3.1-9B 9B Balanced performance and latency
Holo3.1-35B-A3B 35B total / 3B active State-of-the-art production

All models are built on the Qwen 3.5 base and released under Apache 2.0.

💡
Getting started: Download the Q4 GGUF or NVFP4 checkpoint from Hugging Face. Start with the 4B Q4-GGUF model on your own harness and benchmark three metrics: task success rate, latency, and memory. Validate end-to-end function-call parameter mapping before scaling up to the 35B model.

Why Local Inference Matters Now

The timing of Holo3.1 is no coincidence. A widely shared Hacker News post (716 upvotes) demonstrated running Gemma 4 on a GPU-free 2016 Intel Xeon using aggressive speculative decoding. The message from the developer community is clear: the local inference era for agentic AI has arrived.

For teams evaluating computer-use vendors, Holo3.1's quantized variants lower the hidden integration cost that often outweighs raw benchmark scores in production decisions. Privacy-sensitive industries — those that cannot stream screenshots to a cloud API — finally have a production-ready option.

H Company is actively developing a desktop agent harness based on Holo3.1, with more deployment environments planned. The goal is a universal computer-use agent that operates across environments, integrates into any agent stack, and runs wherever the workflow lives.

💬
Availability: All models are Apache 2.0 licensed and available on Hugging Face. H Company's Inference API provides free-tier cloud access for developers who need a managed endpoint before setting up local infrastructure.
Key Points
  • First computer-use model family to ship FP8, NVFP4, and Q4 GGUF quantized checkpoints for local inference
  • AndroidWorld: 35B-A3B jumps from 67% to 79.3%; 4B/9B models from 58% to 72%
  • NVFP4 cuts average step time from 6.8s to 3.3s (~2×); 140ms on RTX 4090
  • Native OpenAI-compatible function calling for seamless LangGraph, CrewAI, AutoGen integration
  • Four model sizes (0.8B–35B-A3B), Apache 2.0 license, full agent stack runs on 12GB VRAM
🔗
Resources · Official Sources · How to Use
— Holo3.1 Official Blog (H Company)
— Holo3.1 Launch Announcement (Hugging Face Blog)
— Holo3.1 Quickstart Guide (H Company Hub)