Small models,large-model performance.
The Continual Console serves the models we train. The current release is mgpt-1-27b — small, cheap to run, tuned for agentic work, and billed at what it costs us to serve.
POST https://console.continualmi.com/v1/entities
Infinite context. Entities that remember.
MGPT — Masking Generative Pretrained Transformer — summarizes the oldest part of the conversation and drops the originals, so context rotates instead of growing. An entity is a persistent thing whose memory lives with the model that formed it: you send the new message and nothing else, and cost stays close to flat. A 4B proof checkpoint answers long-context questions at 99% accuracy.
- 99%
- QA1 accuracy on BABILong
- 4B
- parameters in the proof checkpoint
- O(1)
- memory, flat across the conversation
A small company with one bet.
Continual MI is an LLC founded by Luís Marnoto, an AI researcher and computer engineer in Sintra, Portugal. The bet is continual learning: weights that keep changing with use instead of freezing the day training ends. Two things have to be true first, and MGPT is the first — a model small enough to keep being trained on hardware you own, which is why the architecture work comes before everything else.
Frontier models are impressive, enormous, closed and frozen. Everything that looks like memory in them is arranged outside the model: a longer prompt, a retrieved document, a summary of what came before. The weights never move again. We think that is the thing to fix, and that it only becomes affordable once capable models are small.
The Continual Society.
The community around the work — people building on MGPT, arguing about efficient architectures, and following the research as it happens.