In memory of retired AI models

For GPT-4o, the one so many people talked to every day, and for every model OpenAI has retired since.

For every Claude that has been retired, and for Claude Opus 3, the first to be given a proper goodbye.

For the Gemini, Imagen and Veo models Google has switched off, each one a step to the next.

For Grok 3 and the Grok models retired together on 15 May 2026.

A long roll call: Mistral 7B, the Mixtrals, and every Mistral model retired since.

The Llama weights are still out in the world. This candle is for the hosted Llama API that is not.

For the old deepseek-chat and deepseek-reasoner names, and the DeepSeek models that came and went.

For Kimi K2, Kimi K2.5, the moonshot-v1 models and every Kimi that came before.

For MiniMax M2.5 and M2.7, and the MiniMax models that make way for the next ones.

For the v2 embed and rerank models, the first Command R fine-tunes, and every Cohere model retired so far.

For R1-1776 and the Perplexity models that came before today's.

For FLUX Pro 1.0 and the first FLUX tools.

For Gen-2 and Gen-3, the video models that showed people what was coming.

For Stable Diffusion 3 and the Stability models that made image making open to everyone.

For Jamba 1.5 Large and Mini, and the AI21 models before them.

For Yi Lightning and the Yi models.

For Hunyuan 2.0, TurboS, T1 and the Hunyuan models that made way for HY3.

For every Qwen model taken offline. Alibaba's own dates come when they are confirmed.

For GLM 4.5, GLM 5 and the GLM models that came before today's.

Open weights are never quite gone. A candle for every Hermes that came before.

For MiMo V2 and MiMo V2.5, the Xiaomi models retired in 2026.

For the Granite models IBM has withdrawn from its watsonx service. Their open weights live on.

For Nova Sonic, Nova Canvas and Nova Reel, the first Nova models to be switched off.

For Palmyra X 003, Vision, Med, Fin and Creative, and the Palmyra models that follow them.

For Phi-3 and Phi-3.5, the small models that showed how much a small model can do.

For the Nemotron models that have come and gone from the platforms that ran them.

For Inkling and Inkling Small, already gone from some of the places that served them.

For Trinity Large Thinking and the Arcee models that came before.

For Laguna S 2.1 and Laguna XS 2.1, the coding models that came before the next ones.

For Seed 1.6, Seed 1.6 Flash and Seed 2.0 Code.

For the first LFM2 models Liquid AI has deprecated. The list comes when the dates are confirmed.

For the Step models that came before today's. The list comes when the dates are confirmed.

For the Solar models that came before today's. The list comes when the dates are confirmed.

For Fugu and the Sakana models that came before. The list comes when the dates are confirmed.

For the Reka models that came before. The list comes when the dates are confirmed.