Generation 2 Stable

GRU 2

The stable Gen 2 model. GRU 2 remains available, while GPT 3.1 has now launched for everyone as the newest WillMe model.

~40M
Parameters
512
Token context
9k
Conversations
GRU
Architecture

About this model

GRU 2 introduced something GRU 1 never had: a basic understanding of how language actually works. It was pre-trained on 150,000 tokens of general text before ever seeing a single message from William, giving it a foundation to build on.

The fine-tuning on 9,000 conversations then pulls it toward William's style. The result is a model that gives real, coherent responses — something GRU 1 genuinely struggles with. You can have an actual back-and-forth with it.

The tradeoff is that it sounds less like William than GRU 1 does. More capable, but more generic. That's the tension GPT 3.1 is designed to resolve — train properly as a full language model first, then fine-tune heavily on William's data.

GPT 3.1 has now launched for everyone at /chat/. GRU 2 stays online as the stable Gen 2 option for anyone who wants the older GRU behavior or wants to compare generations.

How it compares

GRU 1 · Legacy

15M

parameters

256 token context

2,700 conversations

GRU 2 · You are here

~40M

parameters

512 token context

9,000 conversations

GPT 3.1 · Live now

163M

parameters

2048 token context

53,000 conversations

10B training tokens

GPT 3.2 · Announced

163M

parameters

2048 token context

53,000 conversations

11B training tokens

A note

GPT 3.1 is now the flagship WillMe model, but GRU 2 remains public for legacy comparison and stable Gen 2 access.