Skip to main content
NewHolo4 is here. One model for screens, code, and tools.Read the announcement → Holo is a family of agent models. Give Holo4 a toolbox and it clicks and types on a screen, writes and runs code, or calls your APIs and MCP tools, often in the same run. The API is OpenAI-compatible: tools in, tool_calls out. Get an API key. The free tier needs no credit card.

​
Models

  • Architecture. holo4-35b-a3b and holo3-1-35b-a3b: MoE, 35B total, 3B active. holo4-27b: dense, 27B. holo3-122b-a10b: MoE, 122B total, 10B active.
  • Limits. Up to 5 images per request (JPEG, PNG, WebP) on every model. Holo3 models cap output at 8,192 tokens. Paid tiers get higher rate limits.
  • Agent format. Every model supports native function calling except holo3-122b-a10b, which uses structured outputs. See Pick an output format.
  • Stable IDs. Switching is a model ID change. Before a model is removed, GET /v1/models sets its deprecation_date and a notice appears here. Pin an ID in production.
Benchmarks and methodology: Holo4, Holo3.1. Holo4 and Holo3.1 are also open weights, including smaller sizes the API does not serve: see Local inference. To try Holo without code, use HoloTab.

​
Data retention

The API logs request time, model, and token counts. It does not store prompts or responses: zero data retention by default. H runs its own models, so it never shares inputs or outputs with third parties. See the privacy policy.

​
Start building

Quickstart

Your first call in a few minutes.

Function calling

The agent loop with tools.

Cookbooks

Runnable browser, code, and MCP agents.

Local inference

Run the open weights yourself.