Generation 3 Training

GPT 3.3

The final Gen 3 model, and a real architectural break from the rest of the lineup: a move from RWKV-style recurrence to a transformer, aimed squarely at the attention performance RWKV kept struggling with.

100.7M
Parameters
2048
Token context
2B
Pretraining tokens
2.6k
Conversations
Transformer
Architecture

GPT 3.3 is in training. These are the final architecture specs — see the Learn pages for what each one means.

What's known so far

Every WillMe model through GPT 3.2 has used a GRU or RWKV-like recurrent architecture — a fixed-size state carried token to token, no attention. GPT 3.3 tests something different: a genuine transformer, trading that fixed-size recurrent state for attention over the full context window. The goal is closing the gap where RWKV specifically underperformed.

It's the last model under the Gen 3 name — an experiment run inside the generation we already know, rather than a leap straight to a new one. If it validates the approach, Gen 4 becomes a full transformer generation, and Gen 5 after that is planned to explore a transformer/RNN hybrid with linear-attention-style mixing.

At 100.7M parameters, GPT 3.3 is noticeably smaller than GPT 3.1/3.2's 163M — and the training data tells a similar story. Pre-training moves to a revised, cleaned, and improved 2B token set — smaller than 3.2's 11B, but higher quality throughout.

Fine-tuning data got the same treatment: the source messages were cleaned and reduced by 84%, and combined with reworked conversation grouping, that leaves 2.6k conversations (from 7.9k messages) to fine-tune on — a 95% reduction from the 53,000-conversation sets behind GPT 3.1 and 3.2. The bet is that less, cleaner data beats more, noisier data.

A note

GPT 3.3 is in training — no release date yet. GPT 3.2 stays the current flagship until it's ready.