Is ci-proto state of the art?Pending benchmarks.Announced Nov 12, 2026; no independent results yet.
Continuous Intelligence's first family of open time-series foundation models for zero-shot probabilistic forecasting. Five sizes: XS, S, M, L and XL, with XL at 167M parameters. Announced November 12, 2026. Public weights and inference code are released; a technical report is described as coming soon. No hosted API is offered. The lab's waitlist is for later releases, not for this model.
The details so far
- Lab
- Continuous Intelligence
- Announced
- Nov 12, 2026
- API
- Not yet
- Open weights
- Published · Public, ungated Hugging Face repo ContinuousIntelligence/ci-proto, with config.json reporting 167,190,592 parameters, plus model.ckpt and model-long.ckpt. At verification, that was the only public model repo under the Continuous Intelligence Hugging Face org
- Parameters
- Five sizes (XS, S, M, L and XL); XL is 167M
- Type
- Family of time-series foundation models for zero-shot probabilistic forecasting, including multivariate series and known-future covariates. The official launch post names five sizes: ci-proto XS, S, M, L and XL (lab launch post and model card)
- Outputs
- Nine quantiles from 0.1 to 0.9. Official pages describe a context window of more than 2 million time steps; the Hugging Face bundle includes model.ckpt plus an optional long-history checkpoint, model-long.ckpt, used on the first forecast above the standard context length (lab launch post, product page, and model card)
- Architecture
- Official launch post: an attention-free time mixer whose compute grows linearly with history length, rather than a Transformer attention mixer (lab launch post)
- Code
- Inference and packaging at github.com/continuousintelligence/ci, installable as pip install continuousintelligence. A technical report is described as coming soon on the official product page (GitHub, PyPI, and product page)
- License
- ci-proto Code and Model Licence 1.0 for the code and weights: based on Apache License 2.0, with added attribution and a one-time notification to Continuous Intelligence for redistribution or use in externally available products or services; purely internal use is exempt from those added requirements. Commercial and research use are allowed under those terms. Third-party components keep their upstream licences (NOTICE includes Apache 2.0). Bundled example data is CC BY 4.0 (LICENCE and NOTICE)
- Hosted access
- No first-party hosted API is stated. The official pages point to local CPU or GPU inference through the Python package
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: self-reported fev-bench SQL skill score of 51.5 for ci-proto XL, on a table dated 11 November 2026 that compared foundation models and excluded closed API models (lab launch post).
Lab claim: self-reported context window of more than 2 million observations, described as 64 times as long as Timestral's (lab launch post and model card).
Sources
LABIntroducing ci-proto (Continuous Intelligence)continuousintelligence.co.uk/news/ci-proto/ · Nov 12, 2026LABci-proto product page (Continuous Intelligence)continuousintelligence.co.uk/products/REPOContinuousIntelligence/ci-proto on Hugging Facehuggingface.co/ContinuousIntelligence/ci-protoREPOcontinuousintelligence/ci on GitHubgithub.com/continuousintelligence/ciLABBen Walker on X: ci-protox.com/benjaminwalker/status/2120864197340955028 · Nov 12, 2026Tracking continues until an independent board lists it.