Introducing PKaI Nano 1.1: A Stronger From-Scratch Model

A New Nano Checkpoint

Today we're introducing PKaI Nano 1.1, our second language model trained from scratch on the PowderKeg Intelligence stack.

Nano 1 gave us our first working foundation. Nano 1.1 is the follow-up: another from-scratch run focused on turning that signal into a stronger base model, a cleaner evaluation story, and a more useful checkpoint for the PKaI model family.

The headline is simple: PKaI Nano 1.1 improves over PKaI Nano 1 on every benchmark in our suite.

What Improved

Across seven accuracy benchmarks, PKaI Nano 1.1 averages 48.83 versus 43.98 for PKaI Nano 1. On the six-task accuracy average used in the Nano 1 release post, Nano 1.1 reaches 51.33 versus 47.42.

The largest visible gain is on LAMBADA OpenAI, where Nano 1.1 moves from 23.33 to 33.82. WikiText perplexity also improves sharply, dropping from 50.13 to 32.33. Lower is better on perplexity, so that is a 35.5% reduction from Nano 1 in this run.

The smaller improvements matter too. WinoGrande only moves from 52.57 to 52.72, but it still moves in the right direction. That kind of across-the-board movement is exactly what we wanted from a second Nano release.

GPT-2 Context

The most direct comparison for this release is Nano 1, but GPT-2 remains a useful public reference point for this model size and evaluation setup.

In the benchmarks listed below, PKaI Nano 1.1 is ahead of GPT-2 Small on every reported metric, including WikiText perplexity. It also reaches into GPT-2 Medium and GPT-2 Large territory on specific tasks: SciQ lands at 80.10, just 0.30 points behind GPT-2 Large, and ARC-Challenge reaches 28.24, ahead of GPT-2 Large and 0.26 points behind GPT-2 XL.

Those comparisons matter because Nano 1.1 is a 194M-parameter Nano-class model, smaller than the GPT-2 Medium, Large, and XL checkpoints referenced on those tasks. Matching or approaching larger public baselines on individual benchmarks is a sign that the stack is moving in the right direction, and it gives us a stronger launch point for the more advanced PKaI releases that come next.

Built From Scratch On Our Stack

PKaI Nano 1.1 was trained from scratch using our own model-training stack.

For us, the release is not only about the checkpoint. It is another validation pass for the training workflow, evaluation process, export path, and the practical systems we use to compare models over time.

That is the pattern we care about: train, measure, learn, improve, and ship the strongest public milestones as the stack gets better.

What This Model Is

PKaI Nano 1.1 is a base language model, not an instruction-tuned assistant. It is designed to continue text, not reliably follow chat-style instructions.

Like other small base models, it can be inconsistent, factually unreliable, and sensitive to prompt formatting. We are releasing it as a research and development milestone: a stronger Nano-family checkpoint, trained from scratch and evaluated transparently against the prior release.

What Comes Next

Nano 1.1 is our most capable model yet, and the strongest release from our training stack so far.

Our research shows there is still room for significant improvement, and more advanced models are already in development with stronger reasoning, broader knowledge, and more capable generation.

Credits and Notes

PKaI Nano 1.1 was trained from scratch by PowderKeg Intelligence.

LAMBADA OpenAI refers to the OpenAI-preprocessed LAMBADA test variant used in our evaluation setup.

GPT-2 Small, Medium, Large, and XL are by OpenAI. Benchmark comparisons are provided for context only; PKaI Nano 1.1 is not affiliated with or endorsed by OpenAI.

Benchmarks are reported from our evaluation setup and should be interpreted as comparative signals rather than absolute claims.