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.
| Skill | Score (one ≈10M checkpoint) | Source |
|---|---|---|
| Reasoning about relationships (family trees) | 88.8% | public dataset (CLUTRR) |
| Tracking a story's world as it changes | 94.1% | 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 answer | 98.0% | our suite |
| Calling a tool with the right arguments | 96.7% | our suite |
| Abstention — declining when unsure (at 50% coverage) | 98.1% | our suite |
| Episodic recall — recognizing something seen before | 100% | our suite |
| Sleep retention — memory recovered after a wipe + consolidation | +34.1 pts | our 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.
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 |
|---|---|
| SmolLM2-135M | 4.7% → 86.3% |
| Qwen2.5-1.5B-Instruct | 22.7% → 90.7% |
| SmolLM2-1.7B | 0% → 89.3% |
| Multi-step planning by search | 93.8% (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.
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.