Home/Performance/State-of-the-art image generation Leonardo models and text-to-speech D

Performance

State-of-the-art image generation Leonardo models and text-to-speech Deepgram models now available in Workers AI

We're expanding Workers AI with new partner models from Leonardo.Ai and Deepgram. Start using state-of-the-art image generation models from Leonardo (Phoenix, Lucid Origin) and real-time TTS and STT models from Deepgram (Nova 3, Aura 1). Build full-stack, low-latency AI applications, all hosted on Cloudflare's global network.

MN

Michelle, NikhilCloudflare

·August 27, 2025

State-of-the-art image generation Leonardo models and text-to-speech Deepgram models now available in Workers AI

GPU-backed inference adds closed-source image and voice models

When Workers AI launched, the bet was that AI models would get both faster and smaller. Cloudflare’s infrastructure was built around that assumption, with specialized GPUs in datacenters worldwide serving inference as close to users as possible. The platform was designed to stay general-purpose, but some workloads fit that hardware especially well: low-latency image generation and real-time audio for voice agents. To push further into those niches, Cloudflare is expanding the Workers AI model catalog to include closed-source partner models, starting with two from Leonardo.Ai and two from Deepgram.

Hosting these models on Workers AI means more than just an inference endpoint. The full Cloudflare Developer Platform can sit around the models: Workers for application logic, R2 for storage, Images for serving and transforming media, plus WebRTC and WebSocket support for realtime voice pipelines. For voice agents, that stack covers speech-to-text, text-to-speech, and turn detection, orchestrated through Cloudflare Realtime.

Leonardo.Ai: Phoenix and Lucid Origin

Leonardo.Ai trains its own generative media models and runs a platform for creating them. Two image models from the company are now available on Workers AI: @cf/leonardo/phoenix-1.0 and @cf/leonardo/lucid-origin.

Phoenix, trained from the ground up by Leonardo, is built for text rendering and prompt coherence. An end-to-end image generation request at 25 steps and 1024x1024 resolution took 4.89 seconds.

curl --request POST \
  --url https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/leonardo/phoenix-1.0 \
  --header 'Authorization: Bearer {TOKEN}' \
  --header 'Content-Type: application/json' \
  --data '{
    "prompt": "A 1950s-style neon diner sign glowing at night that reads '\''OPEN 24 HOURS'\'' with chrome details and vintage typography.",
    "width":1024,
    "height":1024,
    "steps": 25,
    "seed":1,
    "guidance": 4,
    "negative_prompt": "bad image, low quality, signature, overexposed, jpeg artifacts, undefined, unclear, Noisy, grainy, oversaturated, overcontrasted"
}'

Lucid Origin is the newer model in the family, aimed at photorealistic output. At 25 steps and 1024x1024 resolution, an image took 4.38 seconds to generate end-to-end.

curl --request POST \
  --url https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/leonardo/lucid-origin \
  --header 'Authorization: Bearer {TOKEN}' \
  --header 'Content-Type: application/json' \
  --data '{
    "prompt": "A 1950s-style neon diner sign glowing at night that reads '\''OPEN 24 HOURS'\'' with chrome details and vintage typography.",
    "width":1024,
    "height":1024,
    "steps": 25,
    "seed":1,
    "guidance": 4,
    "negative_prompt": "bad image, low quality, signature, overexposed, jpeg artifacts, undefined, unclear, Noisy, grainy, oversaturated, overcontrasted"
}'

Deepgram: Nova 3 and Aura 1

Deepgram develops audio models that let users interact with AI through voice, an interface with higher bandwidth than text thanks to pacing, intonation, and other speech signals. Two of its models are now hosted on Workers AI, and both are designed for very fast inference.

@cf/deepgram/nova-3 handles speech-to-text with high accuracy. @cf/deepgram/aura-1 is a context-aware text-to-speech model that applies natural pacing and expressiveness based on input text. The newer Aura 2 model is expected on Workers AI soon. Alongside these additions, sending binary mp3 files to Workers AI no longer requires converting them into a Uint8 array.

Using the AI binding, a Nova 3 speech-to-text request looks like this:

const URL = "https://www.some-website.com/audio.mp3";
const mp3 = await fetch(URL);
 
const res = await env.AI.run("@cf/deepgram/nova-3", {
    "audio": {
      body: mp3.body,
      contentType: "audio/mpeg"
    },
    "detect_language": true
  });

Via the REST API:

curl --request POST \
  --url 'https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/deepgram/nova-3?detect_language=true' \
  --header 'Authorization: Bearer {TOKEN}' \
  --header 'Content-Type: audio/mpeg' \
  --data-binary @/path/to/audio.mp3

Both Deepgram models also support WebSocket connections, keeping a live bidirectional link to the inference server. Documentation for WebSocket usage with Nova 3 is available in the Developer Docs.

With Cloudflare Realtime, these pieces form a full voice pipeline: