Back home

ci-proto

x+web 0.95
Visit source

Intro

ci-proto is the first time-series foundation model (TSFM) family officially released by the London lab Continuous Intelligence on 12 November 2026. The official blog post Introducing ci-proto, dated 12 November 2026, opens: “ci-proto is out today. It’s the debut model family from Continuous Intelligence and our first step towards a new generation of time series foundation models. It’s available to use now.”

At 12:15 UTC the same day, the company account @Continu_Intel posted that ci-proto is #1 on fev-bench, can recall patterns from more than 2 million observations ago, and is open-weight and available for commercial use. At 13:00 UTC, co-founder and CEO Benjamin Walker posted from @benjaminwalker: “Today we're launching ci-proto, @Continu_Intel's first time-series foundation model,” listing the official claims of a fev-bench lead over Google, Amazon and Mistral, a context window 64 times that of other models, and open weights for commercial use. The lab was founded in London in July 2026 and is one of the teams selected for Stage 1 of SPRIND’s Next Frontier AI Challenge, with up to €3 million in funding.

Officially, ci-proto is a family of five zero-shot forecasting models: after pretraining on many datasets, it can forecast on unseen datasets without task-specific retraining. The largest member, ci-proto XL, has 167 million parameters and is still described as a prototype, but the official blog ranks it #1 among public foundation models on fev-bench (a multivariate forecasting benchmark for TSFMs) with a 51.5% skill score, ahead of TimeH2O-v1 (50.8%), Timestral-209M (50.4%), t0-gamma (49.4%), TimesFM-3 (49.0%) and Chronos-2 (47.3%). Those figures come from the official blog, with a cutoff of 11 November 2026, and the table excludes closed API models; this page does not independently reproduce them. The company also says the five sizes form the Pareto frontier for forecast error against model size and runtime: no other public TSFM, it claims, forecasts more accurately for the same size or speed.

On the method, the blog contrasts Transformer attention used by most existing TSFMs with ci-proto’s “attention-free time mixer”: attention grows quadratically with history length, while the time mixer grows only linearly, which the company says makes much longer context windows practical. Officially, ci-proto can use a context window 64 times as long as Mistral’s Timestral and can draw on patterns from more than 2 million observations ago. A scheduled power-outage example in the blog shows the forecast capturing a drop in electricity use only when an earlier outage is included in a four-year history. To run it, the company provides a Python package, pip install continuousintelligence, for local inference on CPU or GPU; users supply past observations and any known future covariates (such as operating schedules) and receive 9 forecast quantiles per target variable. Weights are on Hugging Face: ContinuousIntelligence/ci-proto, and the code is at github.com/continuousintelligence/ci. The blog says the weights are available for commercial use. The company also has a newsletter and a waitlist for later models.

A note on timing: the company account posted at 12:15 UTC (20:15 Asia/Shanghai), just before the previous 20:32 scan ended; the founder’s launch post was at 13:00 UTC (21:00 Asia/Shanghai), inside this slot. The previous slot did not write the name into the database; this slot records ci-proto against the official blog, official and founder posts, Hugging Face weights and the Python package. Do not treat the company name Continuous Intelligence as the same thing as the model name ci-proto, or treat the official fev-bench ranking as independently audited.

Key points:

Explain

Why it matters

This is an official 12 November 2026 launch of the lab’s first public model family, posted on the company blog, the official X account and the founder’s account, with open weights, a commercial-use statement and an installable Python package. For people tracking time-series foundation models, long-context forecasting and open forecasting tools, it is a verifiable new proper name.

This page only restates the official blog, official posts, the Hugging Face model card and the product site. The fev-bench ranking, 64× context, 2 million observations and “Pareto frontier” claims are the company’s statements and are not independently checked here.

Limits

The official leaderboard excludes closed API models and is dated 11 November; some third-party recaps have used a different rank, so cite the official blog and re-check upstream. Install commands, model sizes and license terms may change. Do not describe ci-proto as a forecasting engine already validated on every industry dataset, or read “available to use now” as a generally available hosted API; current entry points are on the official site.

Usage

  1. Read https://www.continuousintelligence.co.uk/news/ci-proto/ , and compare https://www.continuousintelligence.co.uk/products/ , https://huggingface.co/ContinuousIntelligence/ci-proto , https://github.com/continuousintelligence/ci , and the official posts https://x.com/Continu_Intel/status/2120852944056709139 and https://x.com/benjaminwalker/status/2120864197340955028 ;
  2. To try it, follow the current official instructions (the documented example is pip install continuousintelligence for local inference on CPU or GPU); do not rely only on the commands copied here;
  3. When quoting a #1 fev-bench rank, a 51.5% skill score, 64× context or 2 million observations, name the official blog or official posts and re-check upstream;
  4. Do not describe ci-proto as the company itself, or conflate it with other time-series foundation models such as Timestral, TimesFM or Chronos; re-check official pages before citing this entry.

Advertisement