ci-proto brings local forecasting a long-history alternative to attention
ci-proto brings local forecasting a long-history alternative to attention
Not every useful local model needs to chat. Continuous Intelligence's ci-proto targets numerical forecasts, with a design that substitutes recurrent processing for temporal self-attention.
The November 12 research preprint describes a Mamba-3-based family combining temporal recurrence with attention across variables. It supports probabilistic forecasts and both historical covariates and inputs known ahead of time. The authors say inference cost grows linearly with sequence length and claim leading results for the XL variant on fev-bench's 100 tasks. Those are author-reported findings, not an independently established ranking.
What a local deployment would involve
The official repository documents CPU and GPU operation through the same Python interface. It describes automatic selection of a long-history checkpoint above 16,384 observations, plus nine forecast quantiles. Local bundles are supported, and the documentation recommends pinning a complete Hugging Face commit hash for reproducible downloads.
Its fastest CUDA route has hardware and software requirements, including an NVIDIA Ampere-or-newer GPU and a sufficiently recent driver. CPU compatibility should not be mistaken for a promise that every long-history workload will run quickly on a laptop. We did not install the package or test inference.
Available weights do not mean vanilla Apache terms
The code-and-model licence, dated November 10, explicitly adds conditions to Apache 2.0. Its summary describes public attribution and one-time notification requirements for redistribution and externally available products or services, while exempting purely internal use from those extra requirements. Readers should inspect the actual terms before distribution rather than treating an Apache-derived licence as identical to Apache 2.0.
Guth analysis
The practical opportunity is a narrower kind of AI: predicting a workload, demand curve or sensor series locally instead of asking a general assistant to improvise a forecast. The deployment test is whether it beats a suitable baseline on the user's own series, horizon and hardware. Quantile outputs are useful only if their uncertainty estimates remain calibrated on that data.
Evidence limits: the model-card and company-announcement pages were not directly accessible during this pass. The repository documents weight loading, but we did not verify a download. No independent benchmark reproduction was located. The preprint's submission date is verified; the repository's exact publication/update date is not.
Sources and citations
Revision history
Published .