London lab Continuous Intelligence releases ci-proto forecasting models

Continuous Intelligence says ci-proto XL is the number one foundation model on fev-bench, a benchmark for multivariate forecasting with time series foundation models. Its table gives ci-proto XL an SQL skill score of 51.5 across 100 tasks. On that page, SQL is the name for scaled quantile loss. Closed API models are excluded, and the table is dated 11 November 2026[1]. TimeH2O-v1 follows at 50.8, then Timestral-209M at 50.4, t0-gamma at 49.4, TimesFM-3 at 49.0 and Chronos-2 at 47.3[1]. The company says the five models form the Pareto frontier for forecast error against model size and runtime, and that no other public time series foundation model forecasts more accurately for the same size or speed. It also says ci-proto is available for commercial use[1].

TSFM.ai orders every published FEV Bench result by MASE skill score, pretrained models and task-specific models together. On that list a row named ci-proto is second at 40.80, behind TimeRSI-1.5 at 41.44. Inference time is 0.53 seconds per 100 for ci-proto and 523.95 for TimeRSI-1.5[4]. TSFM.ai says the official interactive leaderboard can change which models appear and can sort by pairwise win rate. The company's first place is an SQL skill score for ci-proto XL, with closed API models excluded[1][4].

The company says ci-proto can recall patterns from more than 2 million observations ago, with a context window 64 times larger than any other time series foundation model's, and 64 times as long as Timestral's, from Mistral[1]. The model card and the repository both put the window above 2 million steps[2][3].

The company says most models in the class use Transformer attention, whose cost grows quadratically with the number of past observations. ci-proto uses an attention-free time mixer, so the computation grows linearly with the length of the history, which the company says makes the longer window practical[1].

The example is a scheduled power outage, labelled as simulated data. With 23 days of history the forecast misses the drop in electricity use. With 4 years of history it captures the drop. Both forecasts receive the same future power outage schedule. The figure is one eight-hour window from the long-context checkpoint[1].

The model card calls ci-proto a pretrained model for zero-shot probabilistic forecasting, developed by Continuous Intelligence Limited[2]. The announcement says it can forecast on unseen datasets straight away, without extra training[1]. Training used a large corpus of real-world and synthetic time series. Evaluation was zero-shot on all 100 fev-bench tasks, with the released preprocessing and calibration in tta.json, scored by scaled quantile loss[2].

Inputs are one or more target series, plus optional past and future covariates, timestamps and related series. The model returns nine quantiles from 0.1 to 0.9, in steps of 0.1, for each target and each step[2][3]. model.ckpt covers histories up to 16,384 steps. Longer histories use model-long.ckpt, and the package selects it automatically[2][3]. The install command is pip install continuousintelligence, for Python 3.10 to 3.14, and the same API runs locally on CPU or GPU[2][3].

The weights, and the code in the repository, are under the ci-proto Code and Model Licence[2][3]. The licence file, version 1.0 and dated 10 November 2026, is based on Apache 2.0 and adds attribution and notification rules. Research and commercial use are free on its summary, subject to those rules. Redistribution, and use in an external product or service, need a public attribution and a one-time notice to Continuous Intelligence. Purely internal use is exempt. It is a custom licence, not a standard open-source licence[3].

Bigger models with new capabilities for real-world data are what the company says comes next[1]. The model card and the repository both link a waitlist, and the repository also links a design partner programme[2][3].

On 20 July 2026 the news list said this London lab had been awarded up to €3 million as one of ten Stage 1 teams in SPRIND's €125 million Next Frontier AI Challenge. The funding, it says, will go to models that learn directly from real-world data as it unfolds in continuous time[5]. The same list says the team co-authors six papers at NeurIPS 2026, including the inaugural company paper Time Flies When You Parallelise, which it says advances the frontier of accurate and computationally efficient time series foundation models[5]. A 25 September 2026 item names Benjamin Walker and says he shared the company's progress at SPRIND's conference in Berlin[5].

The company ranks a 167 million parameter prototype first on its SQL table and says the context window is 64 times longer than other time series foundation models[1]. TSFM.ai's MASE table places ci-proto second, behind TimeRSI-1.5, at a much shorter inference time[4].

The weights run locally on CPU or GPU[2][3]. Commercial use is allowed under the ci-proto Code and Model Licence[1][3]. Larger models are next, with a waitlist and a design partner programme already open[1][2][3].