Introduction

​

Unified Interface for Visual Intelligence

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.vlm.run/v1/openai",
    api_key="sk-...",
)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://gateway.vlm.run/v1/openai",
  apiKey: "sk-...",
});

package main

import (
	"github.com/openai/openai-go"
	"github.com/openai/openai-go/option"
)

func main() {
	client := openai.NewClient(
		option.WithBaseURL("https://gateway.vlm.run/v1/openai/"),
		option.WithAPIKey("sk-..."),
	)
	_ = client
}

// Cargo.toml: async-openai = { version = "0.42", features = ["chat-completion"] }
use async_openai::{config::OpenAIConfig, Client};

fn main() {
    let config = OpenAIConfig::new()
        .with_api_base("https://gateway.vlm.run/v1/openai")
        .with_api_key("sk-...");
    let _client = Client::with_config(config);
}

require "openai"

client = OpenAI::Client.new(
  base_url: "https://gateway.vlm.run/v1/openai",
  api_key: "sk-..."
)

​

Why the Gateway

• Build on open-weight models: No vendor or frontier-model lock-in. Swap open-weight VLMs, ViT models, and other specialized vision models by changing model, and compare output and cost side by side. See Models.
• Our GPUs or yours: Use these models on our GPUs with no hardware to provision, at a published rate on the live model catalog. Or deploy the same models onto your own GPUs with our platform.
• Uncompromised quality: We are not another LLM serving provider. Each model is tuned and checked for visual quality. The aim is the best quality per dollar for vision.
• Orchestration built-in: Worker parallelism, retries, failure handling, GPU utilization, and response healing run on our side, not in your code. Multi-page PDFs are rasterized and fanned out per page, with no splitting or stitching.
• Agent ready: The MCP server exposes file-reading tools to any MCP-aware framework. System One answers typed questions in one read.

​

Next steps