About lab358
We made context durable.
lab358 is an AI research company that turned a new architecture into a product you can deploy — long-context infrastructure that indexes documents once and reuses them, inside your own cloud.
The idea, briefly
Standard transformers re-read every document on every call and cap how much context they can hold. We ship models converted to a sparse, sub-quadratic architecture with that ceiling removed — so context can be indexed once and served efficiently, on far less hardware. The architecture reaches O(n log n) inference by replacing self-attention with a fully differentiable retrieval mechanism, trained jointly under a single cross-entropy loss.
Why conversion, hosted or self-hosted
We chose to convert the open-weight models teams already trust rather than chase a from-scratch frontier model — and to ship the result two ways, so you pick where it runs. Run it managed on lab358 Cloud, a hosted, usage-based workspace where we operate the infrastructure and you're serving in minutes. Or self-host it: we ship the same converted model as an AWS Marketplace container product that deploys into your own Amazon EKS cluster, where everything runs inside your account — your prompts, documents, and outputs never reach us — and billing flows through the invoice your finance team already pays. Same product; your call on the trade-off between speed and control.
Why 358
3 5 8
Three consecutive numbers in the Fibonacci sequence — a pattern of natural, efficient growth. It's the quiet metaphor for a company built to scale context without scaling cost in lockstep. “lab” signals the research rigor behind the architecture.
Boston
Built where deep tech and research talent run deep.
Pre-seed
A small, technical team — credibility from mechanism, not headcount.
Public Benefit Corporation
lab358 is a public benefit corporation, with durable, efficient long-context AI infrastructure as our mission.