TernaT is the first learned neural reasoner that answers complex questions by chaining facts — with a 90% success rate and a 16 KB brain. No LLM. No GPU. No FPU. Fits on a $0.50 chip.
The first learned neural reasoner that works without a language model — delivering multi-hop reasoning at microcontroller cost.
A new approach to on-device AI that answers multi-step questions without a language model or GPU.
We present TernaT, the first learned neural reasoner that uses compact ternary weights {−1, 0, +1} to answer multi-step questions about facts. Our system combines a 16 KB ternary resonator, a FastController (610 KB int8), and a ChainScorer (461 KB int8) to achieve 90% exact multi-hop QA on a benchmark of 96 facts and 30 queries — without any LLM, GPU, or floating-point hardware.
The complete pipeline fits in ~1.1 MB (resonator only 16 KB), deployable on $0.50 microcontrollers for the core reasoning. Key insight: a hybrid exact+VSA memory architecture resolves the accuracy ceiling inherent in pure VSA approaches, while training with counterexamples eliminates incorrect matches in the learned resonator — achieving a 0% false positive rate.
Hybrid exact+VSA memory with learned ternary components — all fitting in a fraction of a modern CPU cache.
Full system achieves 90% exact match — 100% 1-hop, 100% 2-hop, 70% 3-hop. Benchmark: 96 facts, 30 queries.
| Component | Parameters | Size | Deployment |
|---|---|---|---|
| TernMLResonator | 65,536 ternary | 16 KB | Any MCU |
| FastController | 624,264 int8 | ~610 KB | Cortex-M7+ |
| ChainScorer | 471,553 int8 | ~461 KB | Cortex-M7+ |
| VSA Memory (96 facts) | D=1024 | ~44 KB | Cortex-M4+ |
| Total pipeline | ~1.1 MB | $0.50 MCUs |
Four fundamental contributions — resonator fits in 16 KB, full pipeline ~1.1 MB.
First learned neural VSA reasoner — all prior work is algorithmic, not learned.
Combined with TernML perception models (from 15 KB) for complete on-device AI.
TernaT is an independent research project at the intersection of neural reasoning and compact ternary networks.
TernaT proves that learned neural reasoning over facts is not only possible but practical — achieving 90% exact multi-hop QA with a ~1.1 MB pipeline (core resonator 16 KB) deployable on $0.50 microcontrollers. The key insight: a hybrid exact+VSA memory architecture resolves the accuracy ceiling inherent in pure VSA approaches, while training with counterexamples eliminates incorrect matches in the learned resonator — achieving a 0% false positive rate. The NL parser understands plain English, including multi-word entities like "Eiffel Tower".
Combined with TernML for perception (models from 15 KB), TernaT forms the backbone of a complete on-device AI pipeline that requires no LLM, no GPU, and no floating-point hardware.
Publications on Zenodo � open access papers, restricted code repository.
The first learned neural reasoner over Vector-Symbolic Architecture. 90% exact multi-hop QA, 16 KB ternary resonator, no LLM or GPU required.
Download PaperComplete codebase for TernML multi-architecture ternary framework and TernaT neural reasoner. Includes training, inference, and C codegen for Cortex-M0+.
Source CodeInterested in licensing, collaboration, or early access? Reach out directly.
