13B
Unknown
Apache 2.0
Open-weight 13B language model family designed for historical-language and temporal-generalization research. The flagship vintage checkpoints are trained on 260B tokens of pre-1931 English-language text, creating a strict historical knowledge cutoff intended to reduce modern-data contamination and enable controlled experiments on what models can infer without exposure to later events, web text, or code-heavy modern corpora. The family includes a vintage base model, an instruction-tuned vintage post-train, and a matched modern web base model trained on FineWeb with identical architecture and training FLOPs for controlled comparisons between historical and contemporary pretraining data. The instruction-tuned model was fine-tuned on instruction-response pairs extracted from pre-1931 reference works such as etiquette manuals, encyclopedias, and letter-writing manuals, then further improved with online DPO using an LLM-as-a-judge. Release materials position Talkie as a testbed for historical reasoning, style transfer, temporal surprise analysis, and contamination-free evaluation rather than as a frontier general-purpose assistant.