# Talkie

> **TL;DR.** Talkie (full name: talkie-1930) is a 13B open-weight language model trained exclusively on 260 billion tokens of English text published before 1931.

- **Category:** AI / Research / Language Models
- **Stage:** validating
- **Age:** 50 days
- **Origin date:** 2026-04-27
- **First detected:** 2026-04-28
- **Canonical URL:** https://earlyterms.com/term/talkie
- **Sources:** 6 primary URLs

## Definition

Talkie (full name: talkie-1930) is a 13B open-weight language model trained exclusively on 260 billion tokens of English text published before 1931. Its hard knowledge cutoff at December 31, 1930 makes it the largest publicly released "vintage" language model — one trained on a historically bounded corpus rather than the modern web.

Nick Levine, David Duvenaud, and Alec Radford (co-creator of GPT-1, GPT-2, and Whisper) [announced talkie on April 27, 2026](https://talkie-lm.com/introducing-talkie) via a research blog and simultaneous release of two Apache 2.0 checkpoints on Hugging Face: a 13B base model and an instruction-tuned chat variant post-trained entirely on pre-1931 reference works.

## Example

Talkie's researchers fed the model in-context Python examples and asked it to complete HumanEval coding tasks — a language that did not exist in 1930. Talkie produced syntactically correct one-line solutions, demonstrating that core in-context generalization persists even when a model has never encountered the target domain in training.

## Analogy

Think of it as a linguist fluent in 1920s English who learns Python from a crib sheet.

## Why it's emerging now

Benchmark contamination has become one of the most debated problems in LLM evaluation. Talkie — released April 27, 2026 by a team including Alec Radford — offers a structurally contamination-free test environment, and its frontpage HN appearance signals that developer appetite for rigorous, verifiable benchmarking tools is at a peak.

## Related terms

- *related:* vegan model
- *related:* contamination-free benchmark
- *alias:* vintage language model
- *related:* historical LLM
- *related:* llm-wiki
- *related:* deep-research
- *related:* grpo
- *competitor:* qwen3
- *related:* Alec Radford
- *related:* HumanEval
- *parent:* open-weight model

## Sources

1. [talkie-lm — Introducing talkie: a 13B vintage language model from 1930 (official launch post)](https://talkie-lm.com/introducing-talkie)
2. [GitHub — talkie-lm/talkie (inference library, Apache 2.0)](https://github.com/talkie-lm/talkie)
3. [Hacker News — Talkie frontpage thread (490 pts, 191 comments)](https://news.ycombinator.com/item?id=47927903)
4. [Simon Willison — Notes on talkie (Apr 28, 2026)](https://simonwillison.net/2026/Apr/28/talkie/)
5. [MarkTechPost — Meet Talkie-1930: A 13B Open-Weight LLM (Apr 27, 2026)](https://www.marktechpost.com/2026/04/27/meet-talkie-1930-a-13b-open-weight-llm-trained-on-pre-1931-english-text-for-historical-reasoning-and-generalization-research/)
6. [HuggingFace — talkie-lm organization (base + IT models, Apache 2.0)](https://huggingface.co/talkie-lm)

---
_Generated by EarlyTerms · https://earlyterms.com/term/talkie_