Talkie 1930

Talkie 1930 is a 13-billion-parameter open-weight language model trained on 260 billion tokens of pre-1931 English text, drawn from books, newspapers, periodicals, scientific journals, patents, and case law. The instruction-tuned variant (talkie-1930-13b-it) was finetuned on instruction-response pairs extracted from pre-1931 reference works such as etiquette manuals, encyclopedias, and letter-writing guides, then refined with online DPO reinforcement learning using an LLM as a judge to improve instruction-following. Unlike models trained on the modern internet, Talkie 1930 has no knowledge of post-1930 events, technologies, or digital computing. That hard knowledge cutoff makes it useful for computational linguistics research, enabling controlled comparisons between historical and contemporary model behavior without the benchmark data contamination that affects web-trained models. Researchers use it to study generalization beyond pretraining data, including whether a model trained before computers existed can learn modern programming through in-context examples alone, and to measure how surprising post-1930 events are to a model that never saw them. It also offers humanities scholars and artists a distinctive window into pre-1931 knowledge, culture, and worldviews. Released in April 2026 under the Apache 2.0 license, Talkie 1930 is described by its creators as the largest vintage language model they are aware of, with an 8192-token context window and bfloat16 inference requiring roughly 28 GB of VRAM.