Skip to main content
The Models API is OpenAI-compatible: point the official OpenAI client (or any compatible library) at H Company’s endpoint and call chat/completions. Holo-specific behavior, structured outputs, reasoning, and the coordinate convention, is opted into through extra_body and a few conventions documented below.

​
Endpoint

Base URLhttps://api.hcompany.ai/v1/
AuthAuthorization: Bearer $HAI_API_KEY (handled by the OpenAI client)
MethodPOST /chat/completions
KeysCreate one on Portal-H
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.hcompany.ai/v1/",
    api_key=os.environ["HAI_API_KEY"],
)

​
Models

Model IDArchitectureInputOutputContextMax imagesLicense
holo3-1-35b-a3bMoE, 35B / 3B activeText + image$0.25 / 1M65,5365Apache 2.0
holo3-122b-a10bMoE, 122B / 10B activeText + image$0.40 / 1M65,5365Research only
Output is billed at 1.80/1Mtokensfor‘holo3−1−35b−a3b‘and1.80 / 1M tokens for `holo3-1-35b-a3b` and 3.00 / 1M tokens for holo3-122b-a10b. Images accept JPEG, PNG, and WebP. holo3-1-35b-a3b is on the free tier (rate-limited, 10 RPM); holo3-122b-a10b is paid-tier only. holo3-35b-a3b (Holo3) is still served but is being deprecated in favor of holo3-1-35b-a3b on June 15, 2026.

​
Request parameters

Standard OpenAI fields apply (model, messages, temperature, tools, tool_choice, max_tokens). The Holo-specific options live in extra_body:
ParameterWhereTypePurpose
structured_outputs.jsonextra_bodyobject (JSON Schema)Constrain the response, at the decoding level, to a JSON object matching your schema. Use this instead of native tool calls for the structured-output loop.
chat_template_kwargs.enable_thinkingextra_bodybooleanToggle the reasoning channel. true for agent loops, false for single-shot grounding.
reasoning_efforttop level"low" | "medium" | "high"How much the model plans before acting. "medium" is a sensible default for agent loops.
resp = client.chat.completions.create(
    model="holo3-1-35b-a3b",
    messages=messages,
    temperature=0.8,
    reasoning_effort="medium",
    extra_body={
        "structured_outputs": {"json": schema},
        "chat_template_kwargs": {"enable_thinking": True},
    },
)

​
Response

Holo returns two channels on every call:
FieldContents
choices[].message.contentThe action: the structured JSON object (structured-output mode) or the assistant text.
choices[].message.reasoning_contentThe thinking trace. Read it for visibility; do not feed it back into the conversation.
choices[].message.tool_callsPresent only in native function-calling mode (holo3-1-35b-a3b).
reasoning_content is dropped between turns by the chat template Holo inherits. Anything the model must remember has to flow through content. See the Agent loop for how to carry state forward.

​
Conventions

  • Coordinates in [0, 1000]. Holo returns click positions as integers normalized to the image you sent. Scale back to pixels with the image’s own dimensions. Origin is top-left.
  • Image budget. Keep at most the last 3 screenshots in context for best accuracy, even though a request accepts up to 5 images.
  • Output formats. Structured outputs work on both models; native function calling (tools / tool_calls) is holo3-1-35b-a3b only. Pick one and stay in it.

​
Limits and billing

  • Free tier: rate-limited access to holo3-1-35b-a3b (10 RPM), no credit card required.
  • Paid tier: higher rate limits and access to holo3-122b-a10b. Add credits on Portal-H.
  • Billing is per model, per million input and output tokens. By default the API uses zero data retention.
See the Models API pricing and FAQ for current rates, tiers, and billing details.

​
Next steps

Quickstart

Run Holo in five minutes.

Agent loop

How to use Holo in your computer-use harness.

Element localization

Get click coordinates from a screenshot.