Terse Micro is a ternary-weight language model built from scratch — every weight is −1, 0 or +1. Trained end-to-end on 8 billion tokens, it ships as a single ~182 MB file and runs on a CPU — no GPU, no cloud. Apache 2.0.
⚑ Trained end-to-end (8B tokens · SFT · ORPO) · single-file GGUF on Hugging Face · Apache-2.0
~423M params (~320M active). Runs on any phone or CPU. ~182 MB file.
Laptop-class — the next step up the ladder.
Runs on an ordinary 16 GB-RAM laptop — no discrete GPU.
Multimodal, on a gaming laptop — 16 GB RAM + 4 GB GPU.
Frontier-scale on a single server node. Fully multimodal.
Our earlier model — and the groundwork the from-scratch Terse family builds on. A 2B ternary-weight LLM fine-tuned from Microsoft BitNet b1.58 on a single 4 GB laptop; on our own internal, self-run eval it lands third of twelve, ahead of every open-weight system we tested — including Qwen2.5-7B.
Explore Orchid 1.0 →Our own internal, self-run eval (semantic-similarity scoring) — a relative comparison, not an official or standard NLP benchmark.
ternative — the project, and its in-house x86/AVX2 ternary inference engine.
terse — the from-scratch family of ternary-weight models (their own architecture).
orchid — Orchid 1.0, our earlier model: a fine-tune of Microsoft BitNet b1.58.
A family of ternary-weight models built from scratch — small, fast, and private enough to run on the device in your hand. Starting with Terse Micro.
Explore the family →
Model
The first competitive LLM trained and aligned in Colombia. A 2B ternary model, ORPO-aligned, that ranks #3 on our internal benchmark.
Explore the model →The inference engine for ternary-weight LLMs with runtime LoRA — “the llama.cpp of BitNet models.” It serves what no other stack can.
How it works →