Market Briefs / technology

Cloudflare Workers AI retires 24 models on November 20

From November 20, 24 Workers AI models, including Llama 3.1 8B, Llama 3.2 1B and 3B and several DeepSeek, Kimi, GLM and Qwen models, stop responding and return HTTP 410 errors. Cloudflare moved the date from November 15.

Published

What happened?

On November 5 Cloudflare announced it is refreshing the Workers AI catalog. From November 20, 2026 the 24 models listed will stop returning responses; requests fail with HTTP 410 Gone and error code 5028. Cloudflare first said November 15 and then moved the date. Sixteen are language models: deepseek-r1-distill-qwen-32b, nemotron-3-120b-a12b, deepseek-v4-pro-0813, deepseek-v4-flash-0731, qwq-32b, qwen3-30b-a3b-fp8, kimi-k2.6, kimi-k2.7-code, glm-5.2, glm-5.3-flash, llama-2-7b-chat-hf-lora, gemma-2b-it-lora, gemma-7b-it-lora, llama-3.1-8b-instruct-fp8, llama-3.2-1b-instruct and llama-3.2-3b-instruct. The other eight come with suggested replacements: melotts to aura-2-en, whisper-tiny-en to nova-3, four Stable Diffusion and Dreamshaper image models to flux-2-klein-9b, llava-1.5-7b to moondream3.1-9B-A2B, and embeddinggemma-300m to embeddinggemma-2, which was listed as available soon.

Why it matters

Once the date passes, any Worker or REST call that names one of these models fails outright, with no gradual degradation. Chatbots, summarizers, transcription and image features that a freelancer wired up months ago and no longer touch are the likeliest to break unnoticed. Swapping a model is rarely just a string change: prompts tuned for one model behave differently on another, and the image, speech and transcription replacements come from other vendors with their own parameters. Embeddings are the trickiest case, because vectors from different models are not comparable, so changing the embedding model normally means re-indexing the stored data.

What should you do?

Search client repositories, wrangler configs and environment variables for “@cf/” model IDs and match them against Cloudflare’s list, including projects with no recent commits. For each hit, pick a replacement and run a handful of real inputs through both models before switching. Handle error 5028 explicitly so a retired model raises an alert instead of a vague failure. If a client uses embeddinggemma-300m, plan the re-indexing now. Tell affected clients the date and quote the change as a small maintenance job.

Original sources