WillMe GPT 3.2 is live — try it now — Chat with it

Meet William.

An intelligent, monolithic conversational assistant. Built for fast inference, unique personality replication, and seamless daily chatting.

The WillMe Ecosystem

From the very first GRU to the bleeding-edge GPT 3 generation, we are committed to keeping all generations of WillMe alive and accessible.

Released

WillMe GPT 3.1

Gen 3 • RWKV Architecture

Our launched monolithic flagship. Moving away from GRU, GPT 3.1 adopts the highly efficient RWKV architecture to ensure long-range style consistency. General world knowledge, masterfully tuned for chat.

  • • 163M Parameters (+48%)
  • • 2048 Token Context Window (×2)
  • • 10B Tokens (EN & SV)
  • • 53,000 Conversations
  • • Available now at /chat/
Released

WillMe GPT 3.2

Gen 3 • RWKV Architecture

The direct successor to 3.1: the same shape, trained harder. A broader pre-training corpus, plus a targeted patching stage that finds where the model drifts from William's style and corrects it.

  • • 163M Parameters
  • • 2048 Token Context Window
  • • 11B Tokens (EN & SV)
  • • Advanced Patching Fine-Tuning
  • • Available now at /chat/
Training

WillMe GPT 3.3

Gen 3 • Transformer Architecture

The final Gen 3 model, and a real architectural break: the recurrent state gives way to attention over the whole context window. Smaller than 3.1 and 3.2, trained on a heavily cleaned set.

  • • 100.7M Parameters
  • • 2048 Token Context Window
  • • 2B Tokens (Cleaned Set)
  • • Attention, No Recurrent State
  • • Currently in training
Released

WillMe GRU 2

Gen 2 • GRU Architecture

The stable Gen 2 model. GRU 2 stays online for legacy access and comparison, while GPT 3.1 is now the public flagship for most chats.

  • • ~40M Parameters
  • • 512 Token Context Window
  • • Trained on real Discord logs
  • • Available indefinitely alongside Gen 3
Released

WillMe GRU 1

Gen 1 • Original GRU

The model that started it all. Kept purely for legacy access, backwards compatibility, and historical fallback.

  • • 15M Parameters
  • • 256 Token Context Window
  • • Trained on real Discord logs
  • • Available indefinitely alongside Gen 3

What's New

Recent updates

Aug 19, 2026

GPT 3.3 announced

The final Gen 3 model is coming, and it's a real architectural break: a move from RWKV-style recurrence to a transformer, targeting the attention performance RWKV struggled with. 100.7M parameters, 2048 token context, trained on a heavily cleaned dataset. Not in development yet. See what's known →

Aug 12, 2026

GPT 3.2 launched for everyone

WillMe GPT 3.2 is now public at /chat/. It's free to use, runs on Intel Arc hardware, and lives alongside GPT 3.1, GRU 2, and GRU 1 for comparison.

Jul 19, 2026

Site refresh & GPT 3.2 in training

GPT 3.2 training runs are underway. Alongside that: every model page now shares one consistent layout, statuses simplified to Released / Training / Announced, a reworked William page with a field guide to the vocabulary, and an updated FAQ below.

May 28, 2026

GPT 3.2 announced

The next generation is in development. GPT 3.2 builds on 3.1 with 11B training tokens and a new advanced patching fine-tuning stage for more accurate personality replication. It launches Wednesday, 2026-08-12 at 14:00. See the model page →

May 14, 2026

GPT 3.1 launched for everyone

WillMe GPT 3.1 is now public at /chat/. It is free to use, runs on Intel Arc hardware, and lives alongside GRU 1 and GRU 2 for comparison.

Apr 28, 2026

Compare page, mobile chat & rebrand

Launched the model comparison page at /compare. The chat page became fully mobile-compatible, and WillMe AI became a division of Infinity Intelligence under Infinity Productions.

Mar 2026

Site redesign & model overview pages

Moved to a clean light theme with indigo accents. Each model now has its own overview page at /models/gru1 and /models/gru2.

Oct 2025

GRU 2 released

~40M parameter model with 150k tokens of general pre-training before fine-tuning. First WillMe model capable of coherent back-and-forth conversation.

Aug – Sep 2025

GRU 1 — where it started

15M parameters trained exclusively on 2,700 real conversations. Incoherent, chaotic, and somehow very William. Still live.

Frequently asked questions