Ternative is an open-source AI project from Colombia, building competitive language models — and the inference engine to run them — on hardware ordinary people already own.
The dominant story of modern AI is one of scale: more parameters, more GPUs, more capital. Terse is our argument against that being the only path.
Terse is a family of ternary-weight models built from scratch — our own architecture, every internal weight −1, 0 or +1. Terse Micro, the first, folds a whole language model into a single ~182 MB file that runs on a plain CPU. It isn't a frontier model, and we don't pretend it is — but building a coherent ternary model from the ground up, on a weekend's worth of compute, is the harder and more durable breakthrough: the footprint win compounds as the family scales.
It grew out of Orchid 1.0, our earlier 2B model — a fine-tune of Microsoft's BitNet b1.58, aligned with ORPO on a single 4 GB laptop, and the first competitive LLM trained and aligned in Colombia. Orchid proved the idea; Terse rebuilds it from first principles. Serving either one meant solving problems the big labs never hit, so we built the ternative engine — all of it Apache 2.0, with weights, code, and an archived DOI.
The project itself — and its in-house x86/AVX2 inference engine for ternary-weight models.
The from-scratch family of ternary-weight models — their own architecture, starting with Terse Micro.
Orchid 1.0, our earlier model — a fine-tune of Microsoft BitNet b1.58, not an own architecture.
A documented recipe with published failure modes — not a black box. Reproduce the benchmarks with the scripts in the repo.
Inference runs on your machine — CPU-only if needed. No account, no telemetry, no cloud dependency.
Weights, engine, paper and data recipe — all public, free for research and commercial use alike.
Ternative is built and maintained in the open, outside any large lab. Funding goes directly to continued development — better models and a faster engine. If your organization relies on open, reproducible AI, consider supporting it through FLOSS/fund.
Both have a citable record. The model is archived on Zenodo with a permanent DOI.
@software{romerochisco2026ternative,
title = {ternative: Inference Engine for
Ternary-Weight LLMs with Runtime LoRA},
author = {Romero Chisco, Michelangelo},
year = {2026},
license = {Apache-2.0}
}