Skip to content

moche — the omni model small enough to be yours

The omni model small enough to be yours.

One small, unified model that understands and generates — built edge-first for the machine you already own.
No cloud. No queue. No meter running.

status live site blog hf


/// what moche is

The frontier is racing to build giants in datacenters. moche runs the other way — a bet on one small, sovereign model that fits where you already work. Korean-first, bilingual (KO + EN), built from the first tokenizer rule up. Not a demo of a distant future; a thing you download, and it's yours.

/// pre-release — we build in the open. moche is in active development. This org is the live build log: you can watch a from-scratch Korean model being trained tokenizer-first. The omni product ships later; benchmark numbers land when the models are done, not before.

/// architecture — a sovereign core, organs you attach

A single next-token core does the thinking. Every sense and skill — speech, sight, image, motion, memory, tools — is a pluggable adapter you attach or detach. New capability = one file + @register. The core stays sovereign; the ecosystem grows around it.

/// open work

what where status
moche-native — from-scratch KO+EN model, public build log GitHub · 🤗 G1 🔄 pretraining
moche-tokenizer — KO+EN 64k BPE, 5.0× leaner on Korean than GPT-2 🤗 tokenizer ✅ shipped
korean-llm-merging — Korean LLM merging methodology notes GitHub ✅ open
hax-measured-dataset — firsthand local-AI & ai-server benchmark data GitHub · 🤗 ✅ CC BY 4.0
jojangju-KR-31B — Korean multimodal reasoning on gemma-4-31B 🤗 model ✅ released

/// principles

  • Small is the design constraint, not a port. Every stage has one gate: survive Q4 quantization or it doesn't ship.
  • Everything is a vocabulary. Text, speech (codec tokens), image (VQ tokens), tools (function tokens) — one predictor handles all.
  • Baseline first. A trick earns its place only if it holds across size trends. A well-tuned vanilla is the origin of every comparison.
  • Sovereign & honest. Built from scratch, on our own hardware, from public data — and we don't inflate what it can do.

`///` © 2026 MOCHE · BUILT IN A SMALL SERVER ROOM, FOR SMALL MACHINES

Popular repositories Loading

  1. moche-core moche-core Public

    밑바닥부터 만드는 한국어+영어 소형 옴니 모델 — 프롬스크래치 프로젝트 공개 기록

    Python

  2. korean-llm-merging korean-llm-merging Public

    한국어 LLM을 병합(model merging) 방법론 기록

  3. hax-measured-dataset hax-measured-dataset Public

    Firsthand-measured local-AI & ai-server benchmark/ops dataset from Hax (hax.moche.ai). CC BY 4.0, Croissant metadata. Canonical: https://hax.moche.ai/data

  4. .github .github Public

    moche-ai organization profile

  5. jikji jikji Public

    Jikji (직지) — Korean-first personal memory layer for AI agents. One memory, every agent, owned by you.

    JavaScript

Repositories

Showing 5 of 5 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…