Self-hosted GUI-agent research
Run a compact vision-language model behind an internal agent loop while controlling infrastructure and screen data.
Independent tool overview
Holo 2 at a glance
Holo 2 is H Company's downloadable vision-language model family for web, desktop, and mobile computer-use agents. Its 4B and 8B checkpoints are Apache-2.0, while the 30B and 235B models are restricted to noncommercial use without a separate agreement.
Visit the official Holo 2 site ↗
- Product type
- Vision-language models for computer-use agents
- Environments
- Web, desktop, and mobile
- Model sizes
- 4B, 8B, 30B-A3B, and 235B-A22B
- Commercially permissive
- 4B and 8B under Apache-2.0
- Restricted
- 30B and 235B are noncommercial without an agreement
- Current successor
- Holo 3.1
Overview
What Holo 2 is
Holo 2 is a model family for developers building agents that understand a screen, locate interface elements, decide on a next action, and operate across web, desktop, and mobile environments. It is a model component, not a ready-made consumer automation app.
The original release included 4B, 8B, and 30B-A3B variants based on Qwen3-VL. H Company later added a 235B-A22B checkpoint and an agentic localization technique that can refine coordinates over several steps for high-resolution interfaces.
Holo 2 combines UI localization with multi-step navigation. It can be used inside a ReAct-style agent loop, served through vLLM as an OpenAI-compatible endpoint, or integrated through H Company's public cookbook. Thinking mode is enabled by default in the documented vLLM setup.
Licensing depends on the checkpoint. The 4B and 8B weights are Apache-2.0 and can support commercial self-hosting. The 30B-A3B and 235B-A22B repository metadata and release materials identify CC BY-NC 4.0 or research-only, noncommercial terms; commercial users must contact H Company.
Holo 2 weights remain available, but Holo 3 and Holo 3.1 are the current generation. New production projects should benchmark Holo 3.1 first for broader model sizes, local quantizations, current API support, and newer cross-platform robustness.
Use cases
Who Holo 2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
UI element localization
Convert a screenshot and target description into coordinates for clicking buttons, fields, links, and other interface elements.
Cross-platform agent evaluation
Compare the same model family across browser, Ubuntu desktop, and Android automation tasks.
Cost-sensitive prototypes
Test 4B or 8B computer-use models before committing to a larger model or a hosted proprietary agent.
Capabilities
Core Holo 2 features
Cross-platform navigation
Targets browser, desktop, and mobile interfaces rather than being limited to one operating environment.
UI element localization
Predicts click coordinates from screenshots and descriptions of the intended interface element.
Multiple model sizes
Offers compact 4B and 8B variants plus mixture-of-experts 30B-A3B and 235B-A22B checkpoints.
Reasoning mode
Uses generated reasoning tokens to plan navigation and grounding steps; the cookbook documents a dedicated Holo 2 reasoning parser for vLLM.
Agentic localization
Can iteratively refine a predicted location over multiple steps, which H Company says improves difficult high-resolution grounding.
OpenAI-compatible serving
The official vLLM guide exposes Holo 2 through a local chat-completions endpoint for integration with existing clients.
Public weights and cookbook
Hugging Face checkpoints and official notebooks cover local deployment, localization, and API-shaped inference.
Drop-in Qwen3-VL infrastructure
The models inherit Qwen3-VL architectures, allowing use with inference frameworks that support those model families.
Process
How the Holo 2 workflow works
Step 1
Choose the checkpoint and license
Start with 4B or 8B for commercial-friendly experimentation; treat 30B and 235B as noncommercial unless H Company grants written rights.
Step 2
Build a sandboxed test environment
Use test accounts, synthetic data, restricted network access, isolated storage, and reversible tasks before connecting real systems.
Step 3
Serve the model correctly
Follow the H Company vLLM guide, use the required reasoning parser, set image and context limits, and confirm GPU precision support.
Step 4
Define a narrow action schema
Expose only the click, type, scroll, wait, navigate, and answer operations needed for the task instead of arbitrary machine control.
Step 5
Add validation and approvals
Require confirmation before purchases, messages, credential use, data changes, submissions, downloads, or other consequential actions.
Step 6
Evaluate real workflows
Measure end-to-end success, wrong clicks, recovery, retries, latency, token use, and unsafe actions on the exact apps and screen sizes you support.
Step 7
Compare with Holo 3.1
Benchmark the current generation before shipping because newer checkpoints add broader deployment options and current hosted-API support.
Cost
Holo 2 pricing and free plan
Holo 2 weights can be downloaded without a software fee, but licenses and operating costs differ. The 4B and 8B checkpoints use Apache-2.0. The 30B-A3B and 235B-A22B checkpoints are noncommercial without separate permission. Self-hosters pay for GPUs, storage, bandwidth, monitoring, and engineering. H Company's current public API pricing applies to Holo 3.1, not Holo 2.
Holo 2 4B and 8B weights
$0 software license
Downloadable Apache-2.0 checkpoints suitable for research and commercial self-hosting.
- Infrastructure and engineering are separate
- Review the exact repository license before distribution
- No managed service or support is included with the weights
Holo 2 30B-A3B and 235B-A22B
Noncommercial
Downloadable research checkpoints whose public metadata restricts commercial use.
- CC BY-NC 4.0 or research-only repository terms
- Contact H Company for commercial permission
- Larger variants require substantially more inference infrastructure
Commercial support or rights
Custom
Organizations needing restricted checkpoints, enterprise support, or current hosted products should contact H Company.
- Confirm model version and deployment rights
- Confirm support, retention, security, and service levels
- Current hosted offering is centered on Holo 3 and Holo 3.1
Pricing checked . Check current pricing at the source ↗
Assessment
Holo 2 strengths and limitations
Where it stands out
- Public checkpoints cover several performance and infrastructure targets
- 4B and 8B models have permissive Apache-2.0 licenses
- Combines navigation and coordinate localization across three major interface environments
- Official vLLM and OpenAI-compatible examples reduce integration work
- Small variants make controlled local experimentation more practical
- Weights remain available even though H Company has moved to a newer generation
What to consider
- Holo 2 is a previous-generation model family; Holo 3.1 is the stronger starting point for new production work
- It is not a complete agent, browser, sandbox, policy engine, credential manager, or approval system
- Computer-use models can misread screens, click the wrong element, leak information, follow malicious page instructions, or perform unintended actions
- Vendor-reported benchmark results do not guarantee reliability on a specific application, resolution, language, account state, or long workflow
- The 30B and 235B checkpoints are not licensed for ordinary commercial use without an agreement
- Some large-model card text contains inconsistent generic Apache wording even though repository metadata and release materials mark the checkpoints noncommercial; verify rights in writing
- Self-hosting shifts GPU capacity, latency, scaling, patching, monitoring, and incident response to the operator
- Reasoning and repeated screenshot steps can increase latency and inference cost
- A screen-only agent has less reliable state than an API integration and may break when layouts, popups, permissions, or localization change
- Safe deployment requires least privilege, sandboxing, allowlists, audit logs, human approval, and a reliable kill switch
Compare
Holo 2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Agents
Holo3.1
Choose Holo 3.1 for H Company's current computer-use family, including newer small models, local quantizations, stronger cross-platform robustness, and current API support.
Explore Holo3.1 →Agents
Holo 3
Choose Holo 3 when you want H Company's newer 35B or 122B agent models and the documented hosted Models API.
Explore Holo 3 →Agents
Gemini Computer Use
Choose Gemini Computer Use when a managed Google model and API are preferable to self-hosting open weights.
Explore Gemini Computer Use →Questions
Holo 2 FAQs
Is Holo 2 still available?
Yes. The Holo 2 model collection and official cookbook remain available. It is now a previous-generation family, while H Company focuses current hosted products and new releases on Holo 3 and Holo 3.1.
Is Holo 2 open source?
The weights are publicly available, but the licenses differ. Holo 2 4B and 8B use Apache-2.0. The 30B-A3B and 235B-A22B checkpoints are noncommercial without separate permission.
Can Holo 2 control a computer by itself?
No. It supplies visual understanding, localization, reasoning, and action predictions. Developers still need an agent loop, screenshot capture, input controls, sandbox, permissions, validation, approvals, and recovery logic.
Which Holo 2 model should I use?
Start with 4B or 8B for permissive licensing and lower infrastructure requirements. Evaluate larger checkpoints only for research or after obtaining commercial rights, and compare every option with Holo 3.1.
How much does Holo 2 cost?
There is no software charge to download the weights. Self-hosting costs depend on GPU time, storage, bandwidth, engineering, monitoring, and task length. Restricted checkpoints may require a separate commercial agreement.
Does H Company offer a hosted Holo 2 API?
Older cookbook examples show an H Company API pattern, but the current public Models API is centered on Holo 3 and Holo 3.1. Do not treat current Holo 3 pricing as Holo 2 pricing.
What is the difference between localization and navigation?
Localization identifies where an interface element is on the screen. Navigation decides which action should happen next across a multi-step task. Holo 2 was trained for both roles.
Is Holo 2 safe for production automation?
Not by itself. Production use needs sandboxed environments, minimal permissions, approved domains and actions, test accounts, audit logs, human confirmation for consequential steps, and continuous evaluation on the real applications.
Bottom line
Our Holo 2 verdict
Holo 2 remains useful for research and for teams that specifically want a compact, self-hosted computer-use model with Apache-licensed 4B or 8B weights. It should not be the default for a new deployment without comparison: Holo 3.1 is newer, better aligned with H Company's current tooling, and available in more deployment formats. Whatever model wins, the surrounding sandbox, permissions, validation, and approval design will matter more than a leaderboard score.
Visit Holo 2 website ↗Similar Agents Tools
View all Agents AI tools →
Kimi K2 Thinking
Kimi K2 Thinking - Moonshot AI's open-source thinking agent
Microsoft Agent 365
Agent 365 - Microsoft's platform for managing, securing, and governing AI agents, with capabilities like agent registry, performance analytics, and more.
Parallel Search
Parallel Search - Web search tool built from the ground up for AI agents
Lux
Lux - OpenAGI's fast, cost-effective computer-use model that tops industry benchmarks
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.