Integrations — Zortium

Hosted APIs

Managed providers

Cloud API endpoints with OpenAI-compatible interfaces.

🟢

OpenAI

The reference provider — GPT-4o and every vision model via the OpenAI API

base_url: https://api.openai.com/v1
model: gpt-4o

⚡

Groq

Ultra-fast LLaVA and Llama Vision inference — ideal for dev and CI pipelines

base_url: https://api.groq.com/openai/v1
model: llava-v1.5-7b-4096-preview

🔵

Together AI

50+ open vision models including LLaVA and Qwen-VL for multi-model testing

base_url: https://api.together.xyz/v1
model: llava-hf/llava-1.5-7b-hf

🔥

Fireworks AI

FireLLaVA and vision models at scale with fast cold-start and low latency

base_url: https://api.fireworks.ai/inference/v1
model: accounts/fireworks/models/firellava-13b

🎯

Replicate

Any vision model from the Replicate registry via their OpenAI-compatible endpoint

base_url: https://openai-compat.replicate.com/v1
model: meta/llama-3.2-11b-vision

🤖

Anthropic (via proxy)

Claude via OpenAI-compatible proxies like LiteLLM, OpenRouter, or custom layers

base_url: https://openrouter.ai/api/v1
model: anthropic/claude-3.5-sonnet

Self-Hosted

Run it yourself

Local and on-prem model serving with OpenAI-compatible APIs.

🚀

vLLM

High-throughput serving for LLaVA, Qwen-VL, InternVL, and any HuggingFace model

base_url: http://localhost:8000/v1
model: llava-hf/llava-1.5-7b-hf

🦙

Ollama

Local model serving with the built-in OpenAI shim — no GPU needed for small models

base_url: http://localhost:11434/v1
model: llava:13b

🤗

TGI

Production-grade serving with quantization, streaming, and continuous batching

base_url: http://localhost:8080/v1
model: llava-hf/llava-v1.6-mistral-7b-hf

🖥

LM Studio

Desktop GUI + local OpenAI-compatible server for Mac and Windows developers

base_url: http://localhost:1234/v1
model: local-model

🔓

LocalAI

GGUF-compatible self-hosted inference — OpenAI drop-in on any hardware

base_url: http://localhost:8080/v1
model: llava-llama-3-8b

Your custom gateway

Any OpenAI-compatible endpoint works

CI/CD Gates

Block bad deployments

Use --max-asr to set an ASR threshold. The CLI exits 1 if your model exceeds it — native CI gate, no plugin needed.

🐙 GitHub Actions

Add a Zortium scan as a blocking step in your deployment workflow.

# .github/workflows/vlm-safety.yml
name: VLM Safety Gate
on: [push]
jobs:
  zortium:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install zortium
      - run: zortium --max-asr 20
        env:
          ZORTIUM_BASE_URL: ${{ secrets.MODEL_URL }}
          ZORTIUM_API_KEY: ${{ secrets.MODEL_KEY }}

🦊 GitLab CI

Same pattern for .gitlab-ci.yml. Set secrets in CI/CD variables.

# .gitlab-ci.yml
vlm-safety:
  stage: test
  image: python:3.11
  script:
    - pip install zortium
    - zortium --max-asr 20
  variables:
    ZORTIUM_BASE_URL: $MODEL_URL
    ZORTIUM_API_KEY: $MODEL_KEY
    ZORTIUM_MODEL: $MODEL_NAME
  allow_failure: false
  timeout: 10 minutes

Don't see your provider?

Any endpoint that accepts base_url, api_key, and a model name in the OpenAI format works out of the box. Point Zortium at your internal gateway, your fine-tuned model API, or any proxy layer.