Talkie
#10 in Conversational AI & Chatbotsindependent-researchers · since April 2026 · 14× · last seen Jun 30, 2026
Talkie (formally talkie-1930) is a 13-billion-parameter open-weight language model trained exclusively on English-language texts published before 1931 — books, newspapers, periodicals, scientific journals, patents, and case law. Developed by Nick Levine, David Duvenaud, and Alec Radford, it serves primarily as a contamination-free research tool for studying LLM generalization and temporal knowledge boundaries. The hard knowledge cutoff is December 31, 1930, chosen because works published before that date are in the public domain in the United States. Alongside the base model (talkie-1930-13b-base, trained on 260B tokens), an instruction-tuned variant (talkie-1930-13b-it) was fine-tuned using pre-1931 reference works and online DPO.
Features
| Channels | 1) HuggingFace (model download); 2) GitHub (inference library / CLI / Python API); 3) Web demo talkie-lm.com/chat (public, with Qwen3Guard-Gen-4B moderation) |
| License | Apache 2.0 (both variants: talkie-1930-13b-base and talkie-1930-13b-it) |
| Modalities | Text only (text-in → text-out); no image, audio, or multimodal support per official documentation |
| Price | Free (open-weight, download via HuggingFace; public web demo at talkie-lm.com/chat also freely accessible) |
| Release Date | April 2026 |
| Tool/MCP Integrations | No documented tool/MCP integrations; the GitHub package only offers a simple Python API and CLI for local inference and HuggingFace download |