Public company site
Nantiro is building a 1.58-bit ternary AI model family that targets large-model capability with sharply lower memory, energy, and serving cost.
Company
| Word | Meaning | |
|---|---|---|
| N | Neural | Brain-inspired deep learning architecture |
| A | Additive | Replace multiplications with simple additions |
| N | Network | Large-scale transformer network |
| T | Ternary | Weights quantized to {-1, 0, +1} |
| I | Intelligence | Reasoning-capable language models |
| R | Resource | 80% less memory, compute, and energy |
| O | Optimization | Efficiency at every layer of the stack |
The company mission is simple: make powerful AI cheaper to train, cheaper to serve, easier to deploy privately, and easier to measure for energy impact.
Phase-gated 1B, 7B, 30B, 120B, 400B, 800B, and 1T+ model path with 1.58-bit architecture as the core efficiency wedge.
Prepared-data handoff, tokenizer checks, training launch, checkpointing, loss monitoring, recovery, evaluation, and export workflow.
Target users include developers, startups, enterprises, institutions, and teams that need useful AI without frontier-lab infrastructure budgets.
Technology
Dense frontier AI is powerful, but memory bandwidth, GPU supply, electricity, and cooling make it expensive. Nantiro attacks the cost layer directly.
Dense models commonly store weights in FP16/BF16. Nantiro targets ternary weights {-1, 0, +1}, which mathematically need log2(3), or about 1.58 bits per weight. The goal is lower weight memory and fewer expensive multiply-heavy operations at serving time.
For the same parameter count, packed ternary weights can be about 10.1× smaller than FP16 weights. Training still needs high-precision master weights and optimizer state; the largest benefit appears at inference.
Target weight compression
Carbon-linked financing only works after measurement. Nantiro will track tokens, hardware power, energy per useful output, baseline comparison, and third-party MRV readiness from Phase 0 onward.
Measure, report, verify
Training means the model reads tokenized data, predicts the next token, measures error, updates weights, saves checkpoints, and repeats billions of times. That requires H100 GPU hours, fast storage, data preparation, recovery automation, benchmark evaluation, and operators who can keep the run healthy.
Comparison
The public numbers below describe dense AI baselines, not Nantiro operating spend. They are shown to explain the market pain, while our job is to validate a lower-cost path phase by phase.
Contact
GPU partners, early users, supporters, and collaborators can reach us directly.