H Company API
Access H Company instantly with Puter.js, and add AI to any app in a few lines of code without backend or API keys.
// npm install @heyputer/puter.js
import { puter } from '@heyputer/puter.js';
puter.ai.chat("Explain AI like I'm five!", {
model: "holo/holo3-35b-a3b"
}).then(response => {
console.log(response);
});
<html>
<body>
<script src="https://js.puter.com/v2/"></script>
<script>
puter.ai.chat("Explain AI like I'm five!", {
model: "holo/holo3-35b-a3b"
}).then(response => {
console.log(response);
});
</script>
</body>
</html>
List of H Company Models
Holo3 35B A3B
holo/holo3-35b-a3b
Holo3 35B A3B is a vision-language model from H Company built specifically for GUI and computer-use agents, reading screens and executing actions across web, desktop, and mobile interfaces. It's a sparse Mixture-of-Experts model with 35 billion total parameters and 3 billion active per token, fine-tuned from Qwen3.5-35B-A3B. H Company reports it scores 77.8% on OSWorld-Verified and performs well on ScreenSpot-Pro and OSWorld-G, benchmarks for locating and identifying UI elements. Some third-party coverage frames this as beating proprietary models like GPT-5.4 on computer-use tasks at a fraction of the inference cost, given the model's 3B active parameters. Released March 31, 2026 under an Apache 2.0 license, it's aimed at developers building automation agents that click, type, and navigate real interfaces rather than general-purpose chat or writing tasks.
ChatHolo3.1 35B A3B
hcompany/holo3-1-35b-a3b
Holo3.1 35B A3B is an open-weights vision-language model from H Company for computer-use and GUI agents, operating web, desktop, and mobile interfaces from screenshots. It is a sparse Mixture-of-Experts model with 35 billion total parameters and 3 billion active per token, built on the Qwen 3.5 family. It supports native function calling in addition to structured JSON outputs, which suits agent frameworks that return actions as tool calls. H Company reports 79.3% on AndroidWorld, and says it improves on Holo3 on OSWorld and on its Holotab harness. It was evaluated against Holo3, Qwen3.5, Kimi-K2.5, and Claude Sonnet 4.6. Released June 1, 2026 under an Apache 2.0 license, it is aimed at developers building agents that click, type, and navigate real interfaces.
ChatHolo3 122B A10B
hcompany/holo3-122b-a10b
Holo3 122B A10B is a vision-language model from H Company built for GUI and computer-use agents, reading screenshots and deciding the actions needed to complete tasks across web and desktop applications. It is a sparse Mixture-of-Experts model with 122 billion total parameters and 10 billion active per token, based on Qwen3.5. H Company reports a score of 78.85% on OSWorld-Verified, a benchmark of multi-step desktop tasks, compared with 77.8% for its smaller Holo3 35B A3B model. Released March 31, 2026, it is aimed at developers building agents that click, type, and navigate real interfaces, such as browser automation and business software workflows, rather than general chat or writing tasks.
ChatHolo4 35B A3B
hcompany/holo4-35b-a3b
Holo4 35B A3B is a vision-language model from H Company built for computer-use agents. It reads screenshots and produces clicks, typing, code, and tool calls. It is a Mixture-of-Experts model with 35 billion total parameters and 3 billion active per token, based on Alibaba's Qwen3.6-35B-A3B. H Company reports that it improves significantly over its Qwen base models. On the model card, it scores 30.9% on OSWorld 2.0 at $0.61 per task and 34.5% on AutomationBench at $0.02 per task. The larger Holo4 27B variant scores higher on both benchmarks at a higher cost per task on OSWorld 2.0. It supports a 262,144-token context window and is released under the Apache 2.0 license. It is aimed at developers building agents that operate real graphical interfaces.
ChatHolo4 27B
hcompany/holo4-27b
Holo4 27B is a dense 27 billion parameter vision-language model from H Company for computer-use agents. It is built on Qwen3.8-27B and takes screenshots and tool results as input, then returns clicks, typing, code, and MCP or API tool calls across desktop, web, and mobile interfaces. H Company trained it with supervised fine-tuning on 127 billion tokens, mostly agent operation data, followed by reinforcement learning. It scores 85.2% on OSWorld, against 84.3% for Qwen3.8-27B, and 61.7% on OSWorld 2.0, against 48.0% for Qwen3.8-27B. On OSWorld 2.0 it trails Opus 5.5, which scores 81.8%. It also scores 45.4% on AutomationBench. It supports a 262,144 token context window and suits developers building agents that operate real software interfaces.
Frequently Asked Questions
The H Company API gives you access to models for AI chat. Through Puter.js, you can start using H Company models instantly with zero setup or configuration.
Puter.js supports a variety of H Company models, including Holo3 35B A3B, Holo3.1 35B A3B, Holo3 122B A10B, and more. Find all AI models supported by Puter.js in the AI model list.
With the User-Pays model, users cover their own AI costs through their Puter account. This means you can build apps without worrying about infrastructure expenses.
Puter.js is a JavaScript library that provides access to AI, storage, and other cloud services directly from a single API. It handles authentication, infrastructure, and scaling so you can focus on building your app.
Yes — the H Company API through Puter.js works with any JavaScript framework, Node.js, or plain HTML. Just include the library and start building. See the documentation for more details.