Small open models that match the giants.
Continual MI builds efficient models with the MGPT architecture — 4B–9B parameter models tuned to match or beat far larger ones, aimed first at coding agents and served as an API. We sell that API alongside Monte Lua and enterprise agent work to fund the real goal: models small enough to keep learning on their own.
The race is to scale up.
The opening is to do more with less.
Frontier models are impressive, and they are also enormous, closed, and expensive to run. The industry's answer is to keep scaling up. We think the more useful question is how small a model can be and still compete.
So we start with efficiency. MGPT reworks attention through masking to get far more out of the same compute and memory, pushing 4B–9B parameter models toward the capability of models many times their size.
We prove it on ordinary work first — code. MGPT research goes into coding agents: small models trained to match or exceed larger ones, served as an API. Monte Lua, that API, and enterprise agent services fund the research, and the research has one destination: continual learning, models that keep adapting with use. Being small is the prerequisite — weights can only keep changing when the model is light enough to run and tune without a datacenter. So we make small models excellent first, then make them learn. That goal is where the name comes from.
How the pieces line up.
- MGPT · The architecture
Small models, larger-model capacity.
MGPT reworks attention through masking so a model does far more with less compute and smaller weights. The target is 4B–9B parameter open models that match or exceed models many times their size.
- MGPT platform · Coding agents
Prove it on ordinary work.
MGPT research goes into conventional tasks, coding agents first: small models trained with the MGPT architecture and related work to rival much larger ones on code — then offered as an API on the platform.
- Monte Lua · The game
A game that funds the work.
Monte Lua is an endless visual novel — a persistent world generated as you play, built on our in-house engine. It is a product we sell: the revenue helps fund the research, and the game shows the same small models doing real creative work.
- Enterprise agents · Services
Agents doing real company work.
We build custom AI agents inside other companies. It is the near-term cash layer that funds the mission — and the same agentic work the MGPT models are being trained for.
- The end goal · Continual learning
Weights that keep changing.
Once small models reach peak capability, we build continual learning into them — weights that keep adapting with use. Small is the prerequisite: that only works off the datacenter. Reaching it is what the company is named for.
The bet is simple: stop waiting for bigger closed models. Build efficient open ones small enough to own and good enough to rival the giants — aimed at real work, in code first.
Monte Lua, the MGPT API, and enterprise agent services fund one piece of research: making those small models keep learning with use. Small is the prerequisite — continual learning only works off the datacenter — and reaching it is what the company is named for.
Three surfaces, one direction.
Mask-Generative Pretrained Transformer — efficient attention behind small 4B–9B models trained to rival larger ones. First MGPT models coming soon; the platform API serves 10+ models today.
An endless visual novel — a persistent world generated as you play. The product we sell to help fund the research.
Custom AI agents built into other companies — the services layer that funds the mission today.
Society is where the progress, the builds, and the open-model discussion live. The conversation happens on Discord.