Small models with large-model performance.
MGPT is our efficient model architecture. The first MGPT-trained models are coming soon — meanwhile the platform API serves 10+ models today for agentic work, compatible with the pi.dev agent harness, on one Continual subscription and credit balance shared across every product.
The model governs its own memory.
A standard transformer either grows its context every turn or drops old turns and loses the thread. MGPT does neither. The model emits explicit <|mask|> tokens to hide stale messages from direct attention while keeping their compressed traces in the KV cache.
The effective context contracts as the model masks, but the traces left behind behave like a fading-but-durable recollection — enough to answer questions about information it chose to hide turns earlier, at a fixed cost.
Capability without the context bill.
On the BABILong long-context QA1 task, a 4B-parameter MGPT reaches 99% token accuracy with constant memory — outperforming much larger systems that have to keep growing their context to keep up.
Read the benchmark write-upEfficient enough to put your best models everywhere.
Small models, large-model performance.
MGPT — Mask-Generative Pretrained Transformer — is our efficient architecture: open base models finetuned on our growing MGPT masking dataset to match or exceed far larger ones. The first MGPT-trained models are coming soon.
Built for agentic work.
Heavy tool calling, long trajectories, growing context — coding agents and beyond. Compatible with the pi.dev agent harness; point your agents at the platform through a standard OpenAI-compatible endpoint and keep your existing tooling.
One account, one balance.
A single Continual subscription and credit balance works across every product — the MGPT API and Monte Lua. No per-feature buckets to juggle.
The MGPT API
The platform API is live today. Create scoped platform keys, call the OpenAI-compatible endpoint, and route agentic model work through Continual MI — billed against your shared Continual credits. MGPT-trained models arrive on the same endpoint; until then we serve leading open and frontier models.
The conversation around efficient architecture, local open models, and the path toward continual learning lives in the Continual Society.