Workers AI has grown to almost 40 models with the addition of 17 new entries spanning text generation, summarization, text-to-image, and image-to-text tasks. The update follows the eight models added on February 6th, 2024, and arrives alongside a redesigned developer documentation site built for model search and discovery.
Text generation
@cf/deepseek-ai/deepseek-math-7b-instruct@cf/openchat/openchat-3.5-0106@cf/microsoft/phi-2@cf/tinyllama/tinyllama-1.1b-chat-v1.0@cf/thebloke/discolm-german-7b-v1-awq@cf/qwen/qwen1.5-0.5b-chat@cf/qwen/qwen1.5-1.8b-chat@cf/qwen/qwen1.5-7b-chat-awq@cf/qwen/qwen1.5-14b-chat-awq@cf/tiiuae/falcon-7b-instruct@cf/defog/sqlcoder-7b-2
Summarization
@cf/facebook/bart-large-cnn
Text-to-image
@cf/lykon/dreamshaper-8-lcm@cf/runwayml/stable-diffusion-v1-5-inpainting@cf/runwayml/stable-diffusion-v1-5-img2img@cf/bytedance/stable-diffusion-xl-lightning
Image-to-text
@cf/unum/uform-gen2-qwen-500m
Model choices: LLMs, SLMs, fine-tunes, and quantizations
Most LLMs generalize across tasks, but models tailored to a particular use case carry real benefits, so the catalog now mixes large language models, small language models, multilingual support, and fine-tuned or quantized variants.
Among the LLM additions is falcon-7b-instruct, notable for its multi-query attention approach to producing high-precision responses. Multilingual coverage comes from discolm_german_7b, trained for German, and the qwen1.5 family, trained on multilingual data with strong output in both English and Chinese. Qwen spans 0.5B to 14B parameters and showed particularly impressive accuracy in testing. The SLM side includes tinyllama-1.1b-chat-v1.0, a 1.1B parameter Llama variant, and Microsoft's 1.3B parameter phi-2; smaller models are gaining popularity for faster, cheaper inference without lost accuracy. A fine-tuned entry, openchat-3.5, applies Conditioned Reinforcement Learning Fine-Tuning (C-RLFT), a technique that supports open-source language model development using easily collectable mixed-quality data.
Stable Diffusion beyond text prompts
The image side adds new Stable Diffusion pipelines and optimizations. Stable Diffusion XL Lightning produces high-quality images in just two inference steps, and the release opens up input modalities beyond a plain text prompt.
Image-to-image lets developers transform an input image, for example asking Stable Diffusion for a cartoon version of a portrait. Inpainting turns an uploaded image into something new, such as expanding a photo's background or colorizing a black-and-white photo. An inpainting request takes three inputs:
- An image — the original picture to be modified.
- A mask — a monochrome screen highlighting the region to be painted over.
- A prompt — what the model should generate in that space.
The example below shows the inputs and the request template for an inpainting call.

import { Ai } from '@cloudflare/ai';
export default {
async fetch(request, env) {
const formData = await request.formData();
const prompt = formData.get("prompt")
const imageFile = formData.get("image")
const maskFile = formData.get("mask")
const imageArrayBuffer = await imageFile.arrayBuffer();
const maskArrayBuffer = await maskFile.arrayBuffer();
const ai = new Ai(env.AI);
const inputs = {
prompt,
image: [...new Uint8Array(imageArrayBuffer)],
mask: [...new Uint8Array(maskArrayBuffer)],
strength: 0.8, // Adjust the strength of the transformation
num_steps: 10, // Number of inference steps for the diffusion process
};
const response = await ai.run("@cf/runwayml/stable-diffusion-v1-5-inpainting", inputs);
return new Response(response, {
headers: {
"content-type": "image/png",
},
});
}
}
Specialized models
Several additions target narrow tasks rather than general chat: deepseek-math-7b-instruct for math problems, sqlcoder-7b-2 for SQL generation, bart-large-cnn for summarization, and uform-gen2-qwen-500m for image captioning. These are public now, with more demos and tutorial content planned over the following weeks.
Optimizing the catalog
Fast, efficient inference depends on the right serialization format. ONNX is a generally applicable option with broad hardware and model architecture support; converting an ONNX model into a TensorRT engine, a format designed specifically for Nvidia GPUs, can yield faster inference latency and higher total throughput from LLMs. The choice usually hinges on what a given model architecture and the available hardware support best. The new Stable Diffusion pipelines, which chain a series of models for a specific task, use both TensorRT and ONNX.
A redesigned model catalog
The new developer docs list every model on a single page for searchability, with task type shown on the right-hand side, and individual model pages carry code examples. Model details include prompt templates and properties such as context token limits.




