What happened
Continuous Intelligence announced ci-proto on 12 November 2026 as its debut model family and first step toward a new generation of time-series foundation models. The company describes five models, with ci-proto XL the largest at 167 million parameters. It calls the release a prototype and says the models are available to use now, including for commercial use.
The intended use is forecasting across time-series data such as machine sensors, trading feeds, electricity grids and patient vitals. Continuous Intelligence says ci-proto can forecast on unseen datasets without extra training, which it calls zero-shot forecasting. Its explanation contrasts this with statistical approaches such as ARIMA and VAR, which must be fitted separately to datasets, and machine-learning models that often need bespoke training for each new application. The company presents pretraining across many datasets as the route to reuse across domains.
For inference, users provide past observations and any known future inputs, such as operating schedules. The model returns nine forecast quantiles for each target variable, giving a range of possible outcomes. Continuous Intelligence says its Python package can run locally on CPU or GPU and directs users to install it with pip install continuousintelligence.
The release page reports ci-proto XL at number one on fev-bench, a multivariate forecasting benchmark. Its table, dated 11 November 2026, gives XL a skill score of 51.5 across 100 tasks; the listed comparisons are TimeH2O-v1 at 50.8, Timestral-209M at 50.4, t0-gamma at 49.4, TimesFM-3 at 49.0 and Chronos-2 at 47.3. The page says closed API models are excluded. The company also claims the five-model family defines a Pareto frontier for forecast error against model size and runtime.
A central architectural claim is that ci-proto uses an attention-free time mixer. The company says computation for reading history grows linearly with context length, rather than quadratically as in Transformer attention, making longer histories practical. It says ci-proto can use a context window 64 times as long as Timestral's and recall patterns more than two million observations back. An example on the page uses simulated electricity data: four years of history put an earlier scheduled outage in context, and the forecast captures a subsequent drop in electricity use.
What the archive adds
New Runtime's earlier coverage includes Mistral's Timestral time-series model family (2026-10-23) and IBM's Granite-TTM-R3 (2026-11-08). ci-proto is a new release in a known topic: the growing set of time-series foundation models, with a distinct claim around long context and local execution.
Trend signal
This supports the pattern that agent economics is moving toward completed work only weakly: the release emphasizes forecast quality against model size and runtime, but does not report cost per verified task or operational savings. The more immediate signal is that model comparisons increasingly need to account for runtime and context capacity alongside error rates.
Why now
Continuous Intelligence presents this as its first public model and says it plans to move toward larger models with new capabilities for real-world data. The announcement positions current benchmark performance as a starting point, rather than a finished product line.
Why it matters
Forecasting often sits upstream of decisions about maintenance, inventory, energy use and risk. A model that can be tested on local hardware and applied to unseen datasets could reduce the need to build a bespoke forecasting model for every dataset. Quantile outputs may also help teams reason about a range of outcomes instead of treating a single point estimate as certainty.
The release is not yet evidence that ci-proto will outperform established methods in a particular business. Benchmark rank, model size and context length do not settle practical questions such as latency, memory use, calibration, performance under missing or shifting data, or the cost of operating a forecasting pipeline. Teams should treat the model as a candidate for controlled evaluation: compare it with their current baseline on held-out periods, and measure error, runtime and resource use together.
Limits and uncertainty
The benchmark scores and claims about Pareto performance and context length come from Continuous Intelligence's announcement. The page identifies the fev-bench table date and says closed API models are excluded, but these results are not independently verified here. The long-context illustration uses simulated electricity data, so it does not establish gains on production grids or other real-world datasets. The announcement does not quantify hardware requirements, inference latency, calibration, or commercial terms beyond stating that commercial use is available.
Skill to build now
Build a small evaluation harness for one forecasting task your team already runs. Keep a time-based holdout, compare ci-proto with the existing baseline, and record forecast error, runtime and hardware use for each. Include known future inputs such as schedules only when they would genuinely be available at prediction time; otherwise the comparison can overstate usefulness.
Reader takeaway
ci-proto is a commercially usable, locally runnable time-series model family with an attention-free architecture and promising vendor-reported benchmark results. Its long-context and accuracy claims are reasons to test it, not reasons to replace a working forecast before checking performance on your own data.