Model Benchmarks

By the numbers.

How GRU 1, GRU 2, GPT 3.1, GPT 3.2 compare across size, speed, William resemblance, and usability — plus GPT 3.3, currently training.

GRU 1 · Released
GRU 2 · Released
GPT 3.1 · Released
GPT 3.2 · Released
GPT 3.3 · Training

Scale

Parameters are the weights a model learned in training. Context window is how much of the conversation it can read at once, counted in tokens.

Parameters

GRU 1 15M
GRU 2 40M
GPT 3.1 163M
GPT 3.2 163M
GPT 3.3 100.7M

Context Window

GRU 1 256 tok
GRU 2 512 tok
GPT 3.1 2048 tok
GPT 3.2 2048 tok
GPT 3.3 2048 tok
Reading these numbers

Italic marks a figure that is approximate rather than counted: GRU 2 and GPT 3.2 have no exact parameter count on record.

GPT 3.2 launched Wednesday, 2026-08-12. GPT 3.3 is in training. Its architecture is fixed, but it has no measured speed or rating results yet.

Speed

Tokens per second is steady-state generation once the model is warm. Load is the cold start before it can answer anything.
GRU 1 1,136.58 tokens/sec Load 0.11 s · Response 0.044 s · CPU · float32
GRU 2 13,113.76 tokens/sec Load 0.07 s · Response 0.004 s · CPU · float32
GPT 3.1 47.86 tokens/sec Load 3.62 s · Response 1.010 s · Intel XPU · bfloat16
GPT 3.2 36.48 tokens/sec Load 2.22 s · Response 1.206 s · Intel XPU · float32
How this was measured

Cold model loading time and steady-state response generation on our serving hardware. Response results are the mean of three 50-token runs after one warm-up.

Measured August 24, 2026 using the prompt “Tell me about yourself in a few sentences.” Cold load excludes Python module import. TPS counts decoder steps for GRU models and tokenizer output tokens for GPT models. Raw measurements are stored in benchmarks/results/speed/.

Ratings

Scores from people who used the models, averaged over every counted ballot and read live from the vote record.

William Resemblance

Rated 0–100, higher is better. Open “what this measures” for the question people were asked.
GRU 1 46
GRU 2 52.89
GPT 3.1 13.95
GPT 3.2 58.64
What this measures

Does it actually sound like William? Rated 0–100 by everyone who took part.

General Usability

Rated 0–10, higher is better. Open “what this measures” for the question people were asked.
GRU 1 2.13
GRU 2 3.45
GPT 3.1 2.04
GPT 3.2 3.43
What this measures

Is it actually good to talk to? Rated 0–10 across every counted ballot.

Fun

Rated 0–10, higher is better. Open “what this measures” for the question people were asked.
GRU 1 2.89
GRU 2 3.15
GPT 3.1 1.57
GPT 3.2 4.84
What this measures

Is it enjoyable to interact with? Rated 0–10. Fun tends to track novelty and unpredictability — smaller models score higher here.

Overview

Each metric is averaged over every counted ballot, then scaled to 0–100 so the three axes can share one shape. Further from the centre is better.
25 50 75 Resemblance Usability Fun

GRU 1

Resemblance 46.0 · Usability 21.3 · Fun 28.9

GRU 2

Resemblance 52.9 · Usability 34.5 · Fun 31.5

GPT 3.1

Resemblance 14.0 · Usability 20.4 · Fun 15.7

GPT 3.2

Resemblance 58.6 · Usability 34.3 · Fun 48.4

GPT 3.3 · Training

Not plotted — no ratings collected yet

GPT 3.2 leads on resemblance and fun; GRU 2 leads on usability.

Coming next

GPT 3.3 In training. Its scale is listed under Scale above; there are no speed or rating results yet, so it is left out of those sections rather than shown as an empty column.
About these figures

Resemblance, usability, and fun scores are real ratings collected from people involved with the project. Speed results were measured on August 24, 2026 on the serving machine and will vary by hardware and runtime configuration. GPT 3.2 launched Wednesday, 2026-08-12. GPT 3.3 is in training, so only its scale is listed — it has no measured results yet.