One API. Every visual model.

OCR, detection, segmentation, pose, and chat — every visual model behind one OpenAI-compatible endpoint.

Start for FreeView models

Supported providers

  • Google
  • OpenAI
  • Anthropic
  • xAI
  • Meta
  • Moonshot
  • Microsoft
  • Qwen
  • DeepSeek
  • ZAI
  • Baidu
  • Roboflow
  • LightOn
  • RedNote
  • USYD
Explore models

  • One API, Any Visual Model

    Frontier labs and open VLMs behind one OpenAI-compatible endpoint. Swap models without rewriting your client.

  • More Visual Tools

    Detection, segmentation, OCR, pose, and chat — classical CV and VLMs as typed chat completions.

  • Easy to build, Cheaper to Scale

    Drop in the OpenAI SDK, ship on free tier, then scale with per-token and per-image pricing.

Optimize routing for availability, cost, or latency.

Latency-sensitive requests route to the fastest-responding model for the lowest time to first token. Heavier requests fall through to a larger model.

Fastest TTFT right now?
truefalsefalsefalse
  • Selected route

    Kimi K3

    180 ms

  • Gemini 3 Flash

    240 ms

  • Claude Fable 5

    320 ms

  • GPT 5.6 Luna

    410 ms

A vision-only gateway, built for builders.

Generalist gateways route 300 LLMs but barely 5 VLMs, and zero classical CV models. Visual AI is its own stack. We built the OpenAI-compatible router for it.

  • 01

    Detection, OCR, captioning

    Florence-2, DeepSeek-OCR-2, GLM-OCR, dots.mocr, PP-OCRv6, LightOnOCR, SAM 3, RF-DETR, ViTPose, Qwen3-VL 0.8B.

  • 02

    Computer vision as chat completions

    Detection, segmentation, OCR, pose, captioning, region proposal. All return typed JSON. Compose them like tools.

  • 03

    Local-cloud parity via mm-ctx

    Same model ids in pip install mm-ctx and on the gateway. Run on your laptop, ship to prod with the same line of code.

  • 04

    Method routing via extra_body

    One model, many methods. Florence-2 alone exposes caption, od, ocr, region_proposal, and more without juggling separate endpoints.

  • 05

    Structured outputs, by construction

    JSON-schema enforcement at the gateway. CV wrappers emit fixed schemas, VLMs honor response_format.

  • 06

    BYOK and BYOC ready

    Bring your own Replicate, Fal, HF, or Anthropic key. Or run the gateway in your VPC with HIPAA / SOC 2 controls.

10 visual models. One endpoint.

All behind one OpenAI-compatible endpoint with extra_body method routing.

Browse all models

Built for production visual AI.

Frontier labs and open models on one OpenAI-compatible API — with the latency and uptime production visual AI needs.

  • Providers Supported

    14+

  • Added Latency

    <50ms

  • Uptime SLA

    99.9%

  • Models Available

    100+

from openai import OpenAI

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

resp = client.chat.completions.create(
    model="roboflow/rfdetr-nano",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "detect"},
        {"type": "image_url", "image_url": {
            "url": "https://.../street.jpg"
        }},
    ]}],
    extra_body={"method": "detect"},
)

# {"detections": [
#   {"label": "car",    "confidence": 0.92, "bbox_xyxy": [...]},
#   {"label": "person", "confidence": 0.81, "bbox_xyxy": [...]},
# ], "count": 2}
print(resp.choices[0].message.content)

Drop-in OpenAI client. Vision-native.

Point the OpenAI SDK at gateway.vlm.run/v1/openai/, swap the model id, get back typed JSON. Same call signature, every visual model.

Read the DocsStar mm-ctx

Side-by-side

Why a vision-only gateway.

Generalist routers like OpenRouter cover the same dozen frontier labs. We index the long tail of small VLMs, OCR and classical CV models nobody else hosts behind a chat-completions API.

Capabilitygateway.vlm.runOpenRouterReplicateHF InferenceTogether
Small open VLM coverage
10 live
< 3
~5
Partial
< 5
Classical CV as chat completions
—
Partial
—
—
OpenAI-compatible chat
—
Partial
Structured JSON for vision
LLM only
—
—
LLM only
extra_body method routing
—
—
—
—
Local sibling (mm-ctx)
—
—
—
—
BYOK with logging + rate limits
—
—
—
Per-token / per-image pricing
Per token
Per second
Per second
Per token

Pricing calculator

Estimate your document OCR bill

Pick a document type, page volume, and OCR model. See how VLM Run stacks up against native PDF and vision APIs.

Page volume100K PAGES/MO
10K100K1M
Billing period

Input tokens

250M

Output tokens

100M

Estimated cost

$67.50

Cost comparison

dots.mocr
$67.50
Gemini 3.5 Flash
$1.05K
Claude Opus 4.8
$3K
GPT-5.5
$3.5K
Claude Fable 5
$6K

Competitors use the document token count for both input and output. VLM Run uses each model's measured input tokens per page plus the document text as output tokens. Prices from llm-prices.com.

Use cases

What developers ship on the Gateway.

Compose VLMs and CV models like Lego bricks. One key, one bill, one SDK.

Document parsing pipelines

Layout, OCR, and markdown extraction on contracts, statements, and forms with sub-cent per-page economics.

Real-time CV in production

RF-DETR detection plus SAM 3 segmentation under one URL. Sub-100 ms p50 for the detection path.

Multimodal chatbots and agents

Qwen3-VL 0.8B gives you a tiny multimodal chat model with native image support. Compose with detection for grounded answers.

Healthcare imaging triage

BYOK + private VPC, BAA available. Run small VLMs against imaging reports without data leaving your account.

Get an API key. Ship vision today.

Get an API KeyRead the Docs