<tool_call> <function=done> <parameter=result>**Talkie** is a 13B parameter "vintage" language model released in April 2026, designed to simulate the knowledge and linguistic style of a model trained exclusively on English text published before 1931. ### Key Details * **Name**: Talkie (specifically `talkie-1930-13b-base` and instruction-tuned `talkie-1930-13b-it`). * **Training Data**: 260 billion tokens from pre-1931 English-language text (e.g., books, encyclopedias, periodicals). * **Architecture**: 13 billion parameters. * **Purpose**: To study "vintage" LLMs—models frozen in a specific historical era—to analyze how they retain period-specific knowledge, biases, and linguistic styles without modern contamination. ### Capabilities & Behavior Talkie exhibits distinct characteristics of pre-1930 knowledge: * **Historical Knowledge**: It accurately references events and figures up to the late 1920s (e.g., Einstein's relativity, early television demonstrations) but lacks knowledge of post-1930 events (e.g., WWII outcomes, modern technology). * **Linguistic Style**: It generates text in a formal, encyclopedic style typical of early 20th-century reference works, including period-appropriate vocabulary and phrasing. * **Creative Output**: It can write poetry in the Romantic style and respond to queries with the tone of vintage etiquette manuals or letter-writing guides. ### Availability * **GitHub**: [talkie-lm/talkie](https://github.com/talkie-lm/talkie) (reference code and technical details). * **Hugging Face**: * Base model: [talkie-1930-13b-base](https://huggingface.co/talkie-lm/talkie-1930-13b-base) * Instruction-tuned: [talkie-1930-13b-it](https://huggingface.co/talkie-lm/talkie-1930-13b-it) * **Live Demo**: A chat interface is available at [talkie-lm.com](https://talkie-lm.com/), though the live stream may be unavailable in some browsers. ### Research Context Talkie is part of broader research into "vintage" language models, exploring how LLMs can be used for historical simulation, forecasting, and epistemic analysis by restricting their training data to a specific time cutoff.</parameter> </function> </tool_call>