W4M · Research Notes

Research Note · The brain itself · 17 / 19

One small brain, ten kinds of thinking — on a single set of weights

Most of this series has been about wrapping models the world already trained. This note is about the thing doing the wrapping. It is one brain — about ten million parameters, a single checkpoint, no swapping specialists in and out — and here is everything it does today: reason about relationships, track a story's world, do multi-step arithmetic, call tools with the right arguments, decline when it isn't sure, remember what it has seen, and recover what it learned after its memory is wiped.

W4M Research · July 2026 · ~5 min read

1 brain · 10+ skills
A single ≈10M-parameter checkpoint scores 88–100% on seven of eleven reasoning worlds — with deduction, code, and long-context recall lower and reported as measured — one set of weights, no specialist swapped in per world.
100% · +34 pts
Perfect episodic recall: shown something once, it recognizes it again. And after its working memory is wiped, a consolidation cycle recovers +34 points of what it had learned — memory that survives the reset.
22.7 → 90.7
The same brain, attached to a frozen Qwen2.5-1.5B-Instruct base model, lifts it from 22.7% to 90.7% on our reasoning suite — no base weight touched. The edge version of this series' composition thesis.
The short version The reflex answer to "make a model more capable" is "make it bigger." This series keeps testing the other answer. Earlier notes pointed it at frozen datacenter flagships; this one points it at ourselves. The brain behind the whole series is small — about ten million parameters, one set of weights — and yet on a single checkpoint it reasons across ten-plus different worlds, knows when it doesn't know, remembers what it has encountered, and consolidates that memory during a rest cycle so it survives a wipe. It is the edge-native brain the series opened with (Notes 1–6), now measured across its full skill set in one place. Same thesis, turned inward: breadth and judgment don't only come from scale — a small brain, built right, can carry a lot on one set of weights.

A posture note before the numbers, the same one that opens every note in this series. Most of the figures below are self-run on our own evaluation worlds — we say plainly which are built on public datasets and which are ours. These are internal measurements of one checkpoint's profile, not external-benchmark placements, and we mark them as such. We publish because the shape is what matters: one small brain, many kinds of thinking, a single set of weights.

What one brain does today

Every row below is the same brain — one checkpoint, ≈10M parameters — evaluated on a different kind of task. Nothing is fine-tuned per row; the profile is what a single set of weights carries at once.

SkillScore (one ≈10M checkpoint)Source
Reasoning about relationships (family trees)88.8%public dataset (CLUTRR)
Tracking a story's world as it changes94.1%public dataset (bAbI)
Multi-step logical deduction~69%public dataset (ProofWriter)
Multi-step arithmetic (depth 2–4)99.6%our suite
Composition — chaining several learned skills into one answer98.0%our suite
Calling a tool with the right arguments96.7%our suite
Abstention — declining when unsure (at 50% coverage)98.1%our suite
Episodic recall — recognizing something seen before100%our suite
Sleep retention — memory recovered after a wipe + consolidation+34.1 ptsour suite
Long-context recall (to 128k tokens)72.6 / 74.1 / 72.2 / 69.8 (1k / 16k / 128k)public dataset (BABILong)
Code (early — still a frontier for us)~43%our suite

Three of those rows are the ones we find hardest to build with a small model and care about most. Abstention — the brain declining rather than guessing — is the difference between a tool you can trust and one you have to double-check. Episodic recall at 100% means it genuinely recognizes what it has seen, not just pattern-matches near it. And sleep retention is the one that sounds like a metaphor and isn't: wipe the brain's working memory, run a consolidation cycle, and it recovers +34 points of what it had learned — the memory moved somewhere the reset couldn't reach.

One set of weights, evaluated on every world without a specialist swapped in for each. The profile is what the brain carries all at once — not a leaderboard of separate models.

For lineage: the previous flagship, a 6.9M-parameter brain, held a seven-world average of 88.1% and still owns our best marks on deduction (69.8%) and tool-calling (99.6%). The current brain trades a little of that for breadth — more worlds, longer memory, the beginnings of code. The line keeps moving.

The same brain also raises other models

This is where the note rejoins the rest of the series. The brain isn't only a standalone reasoner — attached to a frozen base model, it raises that model on the same reasoning suite, without touching a single base weight. The base models are named below; the point is still the lift, not the label.

Frozen base model (weights untouched)On our reasoning suite: alone → with the brain attached
SmolLM2-135M4.7% → 86.3%
Qwen2.5-1.5B-Instruct22.7% → 90.7%
SmolLM2-1.7B0% → 89.3%
Multi-step planning by search93.8% (best-first search on our internal planning suite; on the external PlanBench benchmark the brain scores 37–44% vs GPT-4's ~34% — see the PlanBench note)
Interface fidelity (what survives the hand-off, in / through / round-trip)99.3 / 96.9–100 / 98.7

Same idea as the flagship notes, one tier down: capability the base model already had, made reachable by a small brain wrapped around it. A frozen model that couldn't do a task alone can do it with the brain attached — and the brain runs at edge scale, so this is a lift you can carry on-device rather than one that queues for a data center.

What we're teaching it next

A small window into where this goes. Today the brain reasons in structured worlds and raises other models; the next thing we're teaching it is English — to read and eventually speak it — so that over time it can stand on its own as something you simply talk to, not only a layer that attaches to another model. The aim is a brain that can do both: attach to a model to raise it up, or run as a small standalone you can converse with. And we're beginning to look beyond text — the same brain, reading from a vision model instead of a language one, as a first step toward images. Early work, reported honestly when there's something real to show, the way everything in this series is.

Honest note Most of the scorecard is self-run on our own evaluation worlds. Four rows build on public datasets — relationships (CLUTRR), world-tracking (bAbI), deduction (ProofWriter) and long-context recall (BABILong) — and the rest are our own constructions. We report all of them as internal measurements of a single checkpoint, not external-benchmark placements. This is a broad but bounded profile: one small brain that does many things competently on one set of weights — not a claim of general intelligence. Each figure is a single skill on a single suite, and the frontier ones (code, at ~43%, most of all) are early and still moving. The attachment figures are measured on the same internal reasoning suite, and no base weight is touched.
For the technical reader One brain, ≈10M parameters, a single set of weights, evaluated across ten-plus task worlds with no per-task fine-tuning — the reported profile is one checkpoint measured on every world. Public-dataset rows: CLUTRR (relationships), bAbI (world-tracking), ProofWriter (deduction), BABILong (long-context, evaluated to 128k tokens); the remaining rows — multi-step math, skill composition, tool-calling with arguments, selective abstention, episodic re-encounter, and post-wipe consolidation ("sleep retention") — are our own constructions. Predecessor: a 6.9M-parameter brain, seven-world mean 88.1%, best-in-line deduction 69.8% and tool-calling 99.6%. Attachment: the brain paired with frozen base models of 135M–1.7B parameters (SmolLM2-135M, Qwen2.5-1.5B-Instruct, SmolLM2-1.7B) raises them on the same internal suite from 4.7/22.7/0% to 86.3/90.7/89.3% with base weights frozen throughout; a search-based planning task reaches 93.8%. All figures are self-run internal measurements; deltas and profiles are the measurement, not external-benchmark standings. The architecture, the training approach, and the method by which the brain attaches to and raises a frozen base model are proprietary — full configs and evaluation logs are available to qualified partners under NDA.

The industry's instinct for "a model that can do more" is "a model that is bigger." This series keeps measuring the other direction. Here it is at its smallest: one ≈10-million-parameter brain, a single set of weights, that reasons across ten-plus worlds, knows when to decline, remembers what it has met, and holds that memory through a reset — and, wrapped around a frozen model, raises it too.

Methodology: all figures are from real, logged runs on a single ≈10M-parameter checkpoint evaluated across every task world without per-task fine-tuning. Rows sourced to public datasets (CLUTRR, bAbI, ProofWriter, BABILong) are run on those datasets; all other rows are our own constructions, and every figure is reported as an internal, self-run measurement rather than an external-benchmark placement. Attachment figures pair the brain with frozen base models (SmolLM2-135M, Qwen2.5-1.5B-Instruct, SmolLM2-1.7B; 135M–1.7B parameters) on the same internal suite, base weights untouched. No result in this note is extrapolated beyond the configurations tested.

Benchmark & attribution: CLUTRR — Sinha et al. (2019); bAbI — Weston et al. (2015); ProofWriter — Tafjord et al. (2020); BABILong — Kuratov et al. (2024). Dataset names belong to their authors; the measurements here are ours.

One small brain — broad, judgment-aware, and it remembers. It attaches to your model, or eventually stands on its own.

A ≈10M-parameter brain that reasons across many worlds on one set of weights, knows when to decline, and consolidates memory that survives a reset — running at edge scale. Qualified partners can request the full evidence: configs, evaluation logs, and the audit ledger.

Request access →