The Ternary Intelligence Stack: An End-to-End Balanced-Ternary Alternative to Binary AI Infrastructure (The death of the bit)
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
We present the Ternary Intelligence Stack (TIS), a vertically integrated toolchain built on balanced ternary arithmetic, in which the 0 state denotes an active hold rather than the absence of a value. TIS provides a ternary-native counterpart at each layer of a conventional AI pipeline: a programming language with three-way pattern matching and an actor model (ternlang); a 51-opcode instruction set synthesised to Verilog-2001 with per-cell clock-gating (BET-ISA v1.0); a stack-based virtual machine also compiled to WebAssembly for in-browser execution (BET VM, via TernStudio); sparse ternary BLAS primitives (cuTern with TSPARSE_MATMUL); an inference primitive that physically skips zero-valued weights rather than masking them (@sparseskip); and a self-evolving language model (albert.). albert. is a [N]-layer dual-stream mixture-of-experts model (≈[N]M parameters, [N] tensors) with 12 cognitive-function experts per layer. It was trained on an eight-language, 177M-token corpus (32k ByteLevel BPE) over approximately 4,250 epochs, with all weights held ternary throughout training and no post-hoc quantization. Rather than fixing depth in advance, the model grows by inserting layers into itself when a training plateau is detected, using Net2Net-style surgery on a Fibonacci-indexed schedule ([5]L → [26]L over the model's lifetime). The residual stream is bifurcated into two 256-dimensional streams, fused at Fibonacci-indexed layers; initialisation is Mandelbrot-seeded to break expert symmetry. We report measurements across sparsity regimes. TSPARSE_MATMUL reduces multiplications by 2.27× at the model's measured 56.2% sparsity, scaling toward an asymptotic ceiling of ≈122× at 99% sparsity. Independently verified AVX2 SIMD kernels show 1.80× speedup at 50% sparsity and 3.0× over INT8 at 90%. On two GPU-free machines, @sparseskip sustains 79.5–92.2 tokens/s at 75% expert skip. The full v3.0 training run cost ≈[€/$ X] in compute under per-second cost tracking and a pause–investigate–resume discipline. The stack is implemented in Rust across nine open-core crates published on crates.io (Patent Pending A50296/2026), developed at RFI-IRFOS, [an independent research institute / a registered FlexCo] in Graz, Austria. This is a living document, maintained alongside active development. Repository: github.com/eriirfos-eng/ternary-intelligence-stack · ternlang.com
Fields of Study & Subjects
Advanced Neural Network Applications
Numerical Methods and Algorithms
Parallel Computing and Optimization Techniques
Publication details
- Publication date
- 1 January 2026
- Language
- EN
- Document type
- Preprint