W4M · Research Notes

Series · The Edge-Native Future · 1 / 6

Does it scale?

Every architecture that isn’t a transformer meets the same question — and most don’t survive it. Here is our curve.

W4M Research · July 2026 · ~5 min read

R² 0.977
a clean power-law fit, where 1.0 is perfect — the field’s own signature of a real, scalable architecture
30M → 1B
a controlled ladder of five model sizes, one identical recipe — every rung on the same power law (962M flagship)
0
training failures — no divergences, no loss-spike rollbacks — across the entire billion-parameter run
The short version As our model trains, its error falls along a smooth, predictable line — so predictable that a curve fit on the first ~3% of training forecast the final result of the entire run to within 0.03%. The same recipe works unchanged at every size we tested, from 30 million to nearly a billion parameters, and the biggest run finished without a single failure. In business terms: scaling this architecture up is a budgeting exercise, not a research gamble — you can price what a bigger model will deliver before spending the money.

A strong result at a few million parameters is easy to set aside. Maybe the model memorized. Maybe it got lucky. Maybe it falls apart the moment you make it bigger. The history of “post-transformer” ideas is mostly a history of things that worked small and broke before they reached the scale where they’d matter. So the honest test of a new architecture is not whether it can solve a hard problem once. It is whether it obeys a scaling law. GoM does — and this note is the curve.

The curve fits its own final loss

We trained a 962-million-parameter GoM language model from scratch. As it learned, its loss fell along a clean power law, fit across more than ten thousand logged points from five million to fifty billion training tokens, with an R² of 0.977. In plain terms: as the model reads more data, its error drops in a smooth, predictable line — the same mathematical signature the major labs use to show that an architecture is real rather than a one-off.

One number is worth slowing down for. Fit the curve on only the early part of training — roughly the first 1.5 billion of the 50 billion tokens, about 3% of the run — and extrapolate more than thirty times further out: it predicts the final loss to within about three hundredths of one percent, 2.7605 forecast against 2.7612 actual. A curve that calls its own endpoint from that far away isn’t noise that happened to trend down. It is an architecture following a law.

The question the field asks every newcomer to prove — does it hold its shape as it grows? — is one we can answer with a curve, not a promise.

The same shape holds across sizes

A single run shows one model improving predictably with data. The harder question is whether the architecture keeps its shape as it gets bigger. So we trained a controlled ladder of GoM models — identical recipe, only the model’s size changed — from 30 million to 962 million parameters, a 32× range, and estimated each model’s loss floor: the best error it settles toward. Plotted against model size, those floors fall along a clean power law with exponent β = 0.245 (R² = 0.938). One architecture, every scale, the same law. That exponent sits inside the 0.07–0.35 band that published scaling studies report for frontier architectures (Kaplan et al. 2020; Hoffmann et al. 2022). Read directly: a bigger GoM is predictably better, and the curve tells you how much the next increment of size buys.

Stable where new architectures usually break

Across the entire billion-parameter run there were zero divergences and zero loss-spike rollbacks. It trained as stably at nearly a billion parameters as at thirty million — which, for a non-attention architecture, is the part that usually doesn’t hold. Most novel architectures don’t fail the idea; they fail the engineering somewhere on the way up. This one didn’t.

The scaling questionTypical new architectureGoM
Predictable loss curveOften noisy or ad hocR² 0.977; fits its own endpoint to ~0.03%
Holds across sizesBreaks before it mattersPower law across 32×, β = 0.245
Trains stably at 1BDivergence, loss spikesZero divergences, zero reverts
Interactive · Watch it
R² 0.977 · 10,190 logged points · 5M→50B tokens
A 962M GoM run fit to a power law (R² 0.977). Extrapolating from early training predicts the final loss at 2.7605; the run landed at 2.7612 — a gap under 0.03%. Points shown are representative of the logged run; the raw per-point log is available to partners under NDA.
For the technical reader Within-run: L(D) = E + A·D−α; the full-run fit gives R² 0.977 over 10,190 logged points (5M → 50B tokens); a separate fit restricted to the first ~1.5B tokens, extrapolated ~30×, predicts the endpoint — 2.7605 vs 2.7612 actual, <0.03%. Cross-size: five rungs (30M / 77M / 143M / 267M / 962M), identical recipe, data mix, and tokenizer; each rung’s floor EN is extracted at a fixed data-exponent α = 0.227 so small- and large-model floors are directly comparable; the capacity law L(N) = B·N−β across the floors gives β = 0.245 (R² = 0.938), inside the published 0.07–0.35 band (Kaplan et al. 2020; Hoffmann et al. 2022). Chart floors are normalized to the 962M floor; the band is ±1 RMSE of fit residuals. Per-rung provenance JSON and the raw point log are available under NDA.
Honest note We are precise about what this is. At this training budget the base cannot yet hold open-ended conversation — its value is as a stable, efficient foundation for trained, task-specific heads, and as the scaling-and-stability result itself. We don’t present it as a chat model. The scaling curve suggests conversation quality is primarily a token-budget question rather than an open research risk — though that budget has not yet been run.

Why a curve is the point

AI’s binding constraint has quietly shifted from whether models can do the work to what the work costs — in compute, in power, in dollars. An architecture that earns more capability per parameter, and keeps earning it as it grows, is a different kind of answer to that problem than building another data center. A scaling law is how you show the lever is real. The rest of this series is what that lever buys: a model that keeps learning after training, reasons where larger systems fall off, and runs on the device in your hand.

Methodology: figures are from a logged 962M-parameter training run (~10,000+ logged points, 5M→50B tokens; power-law fit R² 0.977; fitted vs. actual final loss 2.7605 / 2.7612) and a controlled same-recipe ladder from 30M to 962M parameters (32×), with cross-seed reproducibility validated at the 92M scale (N=3). The cross-size capacity law is fit across a five-rung ladder (30M / 77M / 143M / 267M parameters, anchored by the 962M run) at a fixed data-exponent α = 0.227, giving β = 0.245 (R² = 0.938); each rung's loss floor is a fixed-α extrapolation from its own training trajectory. Zero divergences and zero loss-spike reverts across the billion-parameter run. The base is generation-limited at this token budget and is presented as a foundation for trained heads. Core architecture is a linear-time, non-attention core; construction details are under NDA.

One architecture — from a puzzle solver to a billion parameters.

Next: what that architecture does that a frozen model can’t — it keeps learning while you use it.

w4m.ai
Qualified partners

See the full evidence

The per-rung fit provenance and the raw 10,190-point training log behind this note are available under NDA. Tell us who you are and what you’d like to see.