The Terse family terse

Ternary models,
built from scratch.

Terse is a family of ternary-weight language models — every internal weight is −1, 0 or +1. Own architecture, trained from scratch, in the open, under Apache 2.0. The pitch isn't a leaderboard score — it's the most capability per megabyte and per joule: open, private, and running on the device in your hand.

−10+1 Trained from scratch Pure PyTorch CPU-first Apache 2.0

The family

One family, from a
phone to a datacenter.

Five ternary models on a single ladder — every one ownable, offline, and built on {−1, 0, +1}. Terse Micro is the foundation: the architecture proven in Micro is the base every larger tier inherits. Only Micro exists today — the rest are planned, described here by the hardware they'll run on, not by specs we haven't earned yet.

Terse Micro Available now

~423M params (~320M active). Runs on any phone or CPU. Ships as a ~182 MB file.

text · phone / CPU
Terse Mini Planned

Laptop-class. Built on the Micro base — the next step up the ladder.

text · laptop
Terse Medium-Lite Planned

Runs interactively on an ordinary 16 GB-RAM laptop — no discrete GPU.

text · 16 GB laptop
Terse Medium Planned

Multimodal. Runs on a gaming laptop — 16 GB RAM plus a 4 GB GPU.

text + image · gaming laptop
Terse Pro Planned

Frontier-scale on a single server node. Fully multimodal.

text + image + video · server node

Only Terse Micro exists today. Mini, Medium-Lite, Medium and Pro are planned — described by target hardware and modality; exact sizes and results come when they're trained.


Terse Micro

A whole language model
in a 182 MB file.

Trained end-to-end

A ~423M-parameter mixture-of-experts model, trained from scratch in pure PyTorch on 8 billion tokens, then chat-tuned (SFT) and identity-aligned (ORPO). It's a proof-of-concept and the foundation of the family — the model that validates the architecture the larger tiers inherit. The headline isn't a benchmark; it's size, efficiency, portability and openness.

Footprint
182MB

Ternary weights carry ~1.6 bits of information each; the model ships as one TQ2_0 GGUF (~182 MB, Q6_K embeddings).

Runtime
CPUno GPU

Runs on a plain CPU via the terse-arch llama.cpp fork — no GPU required.

Training
8B tokens

FineWeb-Edu · ~19 tokens/param — Chinchilla-reasonable for its size.

Architecture

Parameters~423M total · ~320M active/token
WeightsTernary — {−1, 0, +1} on internal projections (embeddings, LM head, norms & router stay full precision)
Layers12 · hidden 1024 · FFN 2816
AttentionGQA — 8 query / 2 KV heads · head dim 128 · QK-norm pre-RoPE · RoPE θ=500k
Experts4-expert MoE, top-2 on odd layers · auxiliary-loss-free bias balancing
FFNReLU²-gated
ExtrasMulti-token-prediction head · tied embeddings

Training & runtime

FrameworkPure PyTorch — from scratch, single GPU
Hardware1× NVIDIA RTX A6000 (48 GB) on RunPod
Compute cost~$127 total, end-to-end
Data8B tokens · FineWeb-Edu
TokenizerLlama-3.1 · 128K vocab
Post-training3-epoch SFT → ORPO identity alignment
ServingF32 GGUF → TQ2_0 · llama.cpp fork, branch terse-arch
LicenseApache-2.0

What shipped

One run, three models.

Base

terse-micro-base

The pretrained language model — 8B tokens of FineWeb-Edu. Fluent English continuation; GPT-2-medium-class by design at this data budget.

SFT

terse-micro-sft

Chat model — 3-epoch supervised fine-tune on a 44.5k-conversation corpus. Learns the chat template and turn structure.

Final

terse-micro-orpo

Identity-aligned with ORPO on top of the SFT model. Preference alignment measurably shifts it toward its own charter.


Its character

Honest about being
small.

Terse has a deliberate identity, tuned in during alignment: English-first (bilingual is the goal for later tiers), built to democratize AI — capable, private, local, free of paywalls. Developed in 2026 by Michelangelo Romero Chisco.

It's honest about its limits — it owns being a small model rather than bluffing, stays neutral on contested topics, and keeps your data on your machine. Playful when the moment's casual, precise when there's real work on the table. No stereotypes, no grandstanding.


The honest win

Capability per megabyte, not per parameter

  • ~1.6 bits/weight. Non-embedding weights are ~10× smaller than fp16 — it fits in phone RAM.
  • Ternary matmuls are adds and subtracts — no multiplies — so it's low-energy per token.
  • MoE top-2 keeps only ~320M of 423M parameters active per token.
  • The same architecture keeps the footprint win as the family scales up on more data.
What we're not claiming

Straight about the ceiling

  • No benchmark marketing. Micro is a proof-of-concept at a ~$130 compute budget — roughly GPT-2-medium class, and we won't pretend otherwise.
  • Not a polished assistant. Its identity is correct at the preference level, but the small, data-limited base doesn't always express it fluently.
  • Data, not architecture, is the limit. Micro saw ~500–2000× fewer tokens than peers. The fix — more data — is exactly what the larger tiers add.

In the open

Code and weights,
all public.

micro-terse
The Terse Micro model & training code — pure PyTorch, from scratch
github.com ↗
Download · GGUF (~182 MB)
The single-file TQ2_0 model — hosted on Hugging Face
huggingface.co ↗
llama.cpp · terse-arch
Inference fork that runs the Terse architecture (branch terse-arch)
github.com ↗

Terse Micro is trained — and downloadable now.

The single-file model, code, and inference fork are all public.