Private desktop agents
Run the model and agent loop on local hardware so screenshots and actions stay within the user's network.
Independent tool overview
Holo3.1 at a glance
Holo3.1 is H Company's open family of vision-language models for computer-use agents, spanning 0.8B to 35B-A3B parameters with support for web, desktop, mobile, function calling, cloud APIs, and fully local deployment.
Visit the official Holo3.1 site ↗
- Model sizes
- 0.8B, 4B, 9B, and 35B-A3B
- Environments
- Web, desktop, and mobile
- License
- Apache 2.0
- Local formats
- BF16, FP8, NVFP4, and Q4 GGUF
- API context
- 65,536 tokens for 35B-A3B
Overview
What Holo3.1 is
Holo3.1 is built for agents that operate graphical interfaces by looking at screenshots and deciding what to click, type, or do next. It targets browser automation, desktop software, Android workflows, and multi-application business tasks rather than general chat.
The family includes 0.8B, 4B, 9B, and 35B-A3B models based on Qwen 3.5. The smaller checkpoints trade accuracy for lower hardware cost, while the mixture-of-experts 35B model activates roughly 3B parameters per step and is H Company's strongest open Holo3.1 option.
Holo3.1 can be self-hosted under Apache 2.0 or accessed through H Company's OpenAI-compatible API. Quantized 35B checkpoints in FP8, NVFP4, and Q4 GGUF make private local execution possible on supported NVIDIA and Apple hardware.
Use cases
Who Holo3.1 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Cross-platform automation
Build one computer-use stack for websites, native desktop apps, Android interfaces, and multi-app workflows.
Cost-sensitive agent loops
Use smaller models, quantized weights, or the low-cost hosted API for tasks that require many screenshots and steps.
Custom agent frameworks
Integrate structured JSON outputs or native function calling into an existing automation harness.
Capabilities
Core Holo3.1 features
Cross-environment navigation
Understands and acts across browser, desktop, and mobile graphical interfaces.
Four model sizes
Choose from lightweight 0.8B, 4B, and 9B models or the higher-performing 35B-A3B mixture-of-experts model.
Function calling
Adds native tool-calling protocols alongside Holo's structured JSON action format.
Quantized local checkpoints
The 35B release includes FP8, NVFP4, and Q4 GGUF variants for different accelerators and local runtimes.
OpenAI-compatible API
Use the hosted 35B-A3B model through a familiar text-and-image API without changing the surrounding agent architecture.
HoloDesktop CLI
H Company's open desktop client can connect to the managed API or a locally served Holo3.1 endpoint.
Process
How the Holo3.1 workflow works
Step 1
Choose a model size
Balance task difficulty, latency, memory, privacy, and hardware constraints before selecting a checkpoint.
Step 2
Select hosted or local inference
Use H Company's API for quick testing or serve an Apache-licensed checkpoint with a compatible local runtime.
Step 3
Connect an agent harness
Pair the model with HoloDesktop or another framework that captures screenshots and executes approved actions.
Step 4
Constrain the task
Limit accessible apps, accounts, file paths, and high-impact actions, and require confirmation where appropriate.
Step 5
Evaluate real workflows
Measure success, latency, retries, cost, and failure recovery on the interfaces your organization actually uses.
Cost
Holo3.1 pricing and free plan
The Holo3.1 model weights are free to download under Apache 2.0; self-hosters pay their own hardware costs. H Company's hosted 35B-A3B API includes a free rate-limited tier and usage-based paid access.
Open weights
Free
Download and self-host Holo3.1 under Apache 2.0.
- 0.8B, 4B, 9B, and 35B-A3B checkpoints are available.
- Infrastructure, electricity, and operations are not included.
Hosted free tier
$0
Rate-limited API access to Holo3.1 35B-A3B for testing.
- No credit card required.
- Current limit is 10 requests per minute.
Hosted paid tier
$0.25 input / $1.80 output per 1M tokens
Pay-as-you-go access to the Holo3.1 35B-A3B API with higher rate limits.
- Text and image input; text output.
- Maximum five images per request.
- Credits are generally non-refundable.
Pricing checked . Check current pricing at the source ↗
Assessment
Holo3.1 strengths and limitations
Where it stands out
- Covers web, desktop, and mobile automation in one model family.
- Open Apache 2.0 weights support commercial modification and private deployment.
- Multiple sizes and quantizations provide useful cost, speed, and accuracy tradeoffs.
- Hosted API pricing is low for screenshot-heavy, multi-step agent loops.
- Works with function-calling and structured-action agent frameworks.
What to consider
- Holo3.1 is a model family, not a complete no-code automation product; it needs an execution harness and safety controls.
- The 35B variants still require substantial local memory and compute despite quantization.
- Smaller checkpoints can lose accuracy on complex or unfamiliar interfaces.
- Vendor benchmark results do not replace evaluation on your own apps, permissions, and failure cases.
- Computer-use agents can click destructive controls or expose sensitive data unless access is tightly scoped and consequential actions require approval.
Compare
Holo3.1 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Agents
Gemini Computer Use
For Google's hosted specialized model for agents that interact with interfaces.
Explore Gemini Computer Use →Consumer
Portable Computer
For a packaged local-first agent experience designed around personal NVIDIA hardware.
Explore Portable Computer →Agents
Perplexity Personal Computer
For a consumer-oriented local orchestrator across files, native apps, and the web.
Explore Perplexity Personal Computer →Agents
My Computer
For a desktop product that brings Manus agent workflows onto a user's local machine.
Explore My Computer →Questions
Holo3.1 FAQs
What is Holo3.1?
It is a family of vision-language models from H Company designed to power agents that navigate web, desktop, and mobile interfaces.
Can Holo3.1 run locally?
Yes. H Company publishes open checkpoints, including quantized 35B variants for NVIDIA and Apple hardware, and HoloDesktop can connect to a local OpenAI-compatible server.
Is Holo3.1 open source?
H Company publishes the Holo3.1 model family under the Apache 2.0 license, including the hosted 35B-A3B model's weights.
Which Holo3.1 size should I use?
Use 0.8B, 4B, or 9B when hardware and latency matter most; use 35B-A3B when stronger navigation accuracy justifies more memory and compute.
How much does the Holo3.1 API cost?
The hosted 35B-A3B API has a free 10-RPM tier. Paid access is currently $0.25 per million input tokens and $1.80 per million output tokens.
Does Holo3.1 execute clicks by itself?
No. The model interprets screenshots and produces actions; an agent harness such as HoloDesktop must capture the screen, execute those actions, and enforce permissions.
Bottom line
Our Holo3.1 verdict
Holo3.1 is a compelling foundation for teams that want cross-platform computer use without committing to a closed frontier model. Its open license, small-to-large model range, local quantizations, and inexpensive API make experimentation accessible. Production use still depends on a reliable harness, careful task-specific evaluation, and strict approval boundaries around sensitive or irreversible actions.
Visit Holo3.1 website ↗Similar Agents Tools
View all Agents AI tools →
Gemini Spark
Gemini Spark - Google's personal agent that runs 24/7 on Cloud VMs
Hermes Desktop
Hermes Desktop - Nous Research's persistent-memory agent as a native desktop app
Higgsfield Supercomputer
Higgsfield Supercomputer - ACloud-based AI agent with built-in tools, persistent memory, and scheduled task automation
Kimi Work
Kimi Work - Moonshot's local desktop agent running up to 300 parallel agents
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.