talkie-1930-13b-it
An open-source 13-billion-parameter language model whose entire training corpus is English text published before 1931. No fine-tune. No "act 1930" system prompt. The cadence, the vocabulary, the ignorance — all of it written into the weights.
Talk to the Past is the only iOS app running it.
A language model with a closed corpus.
talkie-1930-13b-it is an open-source large language model with thirteen billion parameters. It was pre-trained from the very first token of its corpus on English text written before the year 1931 — newspapers, novels, scientific papers, parliamentary records, letters, the Encyclopædia Britannica, the Times of London through October 1930.
There was no fine-tune on modern instruction data. There is no "stay in character" system prompt at inference time. The model has never seen a sentence about World War II, the moon landing, the internet, Sputnik, or anything else that hadn't yet happened when the corpus closed.
It was released in April 2026 under the Apache 2.0 licence. Weights, tokenizer, and a paper describing the corpus construction are public on Hugging Face and GitHub.
It cannot drift to a modern register.
Most "historical AI" products wrap a modern model — GPT-4, Claude, Gemini — with a system prompt instructing it to roleplay as a figure from the past. The illusion holds for a few turns and then collapses. Ask any of those models about Wall Street in October 1929 and you'll get a Wikipedia-style summary in modern prose: "The crash, also known as Black Tuesday, was triggered by..."
Ask talkie-1930-13b-it the same thing and it tells you what people in 1930 thought was happening — what the Times wrote in November 1929, what President Hoover said the month after, what the leader-writers were predicting for the spring. Not because it has been instructed to. Because that is what English text written before 1931 contains, and that is the only English text it has ever read.
Modern models can be coaxed into period prose. They cannot be made ignorant of the future. talkie-1930-13b-it is ignorant by construction. The bewilderment, when you ask it about television or the Beatles or computers, is real — those tokens have no neighbours in its weights.
A small team. A serious paper.
The model was released in April 2026 by:
- Alec Radford — the OpenAI researcher who first-authored GPT-2 and Whisper.
- David Duvenaud — University of Toronto Computer Science professor and former Alignment Evals lead at Anthropic.
- Nick Levine — independent researcher and former financial historian at Winton.
The model is licensed under Apache 2.0. Talk to the Past is an independent iOS application; we are not affiliated with the model's authors.
Selected coverage.
-
A write-up of the model's release on the LLM newsletter of record.
-
Front-page discussion thread on the model's release.
-
Article headline: "Vintage chatbot lives in the past — and won't be talked out of it."
-
Article headline: "Talkie is a vintage LLM trained on pre-1930 data to help facilitate time travel."
-
Article headline: "Meet talkie-1930: a 13B open-weight LLM trained on pre-1931 English text for historical reasoning and generalization research."
The only iOS app running it.
The model is too large to run on a phone. Talk to the Past calls it through a small proxy we operate, streams the reply back letter by letter, and stores the conversation only on your device.
From the source.
-
Hugging Face — model card & weights.
huggingface.co/talkie-lm/talkie-1930-13b-it -
GitHub — corpus construction code & paper.
github.com/talkie-lm/talkie -
Launch blog — "Introducing Talkie".
talkie-lm.com/introducing-talkie -
Project homepage.
talkie-lm.com