Continuous Intelligence, a London lab founded in July, released its first time-series foundation model, ci-proto, with weights on Hugging Face and code on GitHub. It forecasts series such as power load, sales or sensor data zero-shot on CPU or GPU. The company says it reads more than 2 million steps of history, 64 times other time-series foundation models, and returns nine quantiles from 0.1 to 0.9. Its results on the 100 fev-bench tasks were merged into the official fev repo by an AutoGluon maintainer: 51.51 SQL skill and 40.80 MASE skill. The licence builds on Apache 2.0 but adds attribution and a one-time notice for external products.
Key Takeaways
- ✓Size: about 167M parameters (config lists 167,190,592), plus a long-history checkpoint; histories up to 16,384 steps use the standard checkpoint and longer ones switch automatically
- ✓Context: the company says it reads 2M+ steps of history, about 64x other time-series foundation models (vendor claim)
- ✓Score: 51.51 SQL skill and 40.80 MASE skill zero-shot on 100 fev-bench tasks, submitted by the vendor and merged on Nov 11 in fev PR #211
- ✓Outputs: joint multivariate forecasts with past and future covariates and categorical inputs, returning nine quantiles from 0.1 to 0.9
- ✓Licence: custom licence based on Apache 2.0, commercial use allowed; external products or redistribution need public attribution and a one-time notice, internal use is exempt
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain points
Grid load, retail sales and equipment sensors are all time series. Teams used to tune a statistical model per series; recent time-series foundation models such as Google TimesFM, Amazon Chronos and Mistral Timestral forecast zero-shot instead, but usually see only thousands to tens of thousands of points, so yearly seasonality and rare spikes get cut off. Continuous Intelligence, a London lab founded in July, pitches its first model, ci-proto, on longer memory.
Architecture highlights
The public ci-proto has about 167M parameters and ships a standard and a long-history checkpoint: histories up to 16,384 steps use the standard one, longer ones switch automatically. The company says it reads 2M+ steps, about 64x comparable models. Input is a variables-by-time tensor forecast jointly by default, with past and future covariates, categorical inputs and missing values, and output is nine quantiles from 0.1 to 0.9. The README says Linux x86-64 installs a fused Mamba-3 CUDA backend, with a CPU fallback.
Benchmarks and data
fev-bench is the AutoGluon team's 100-task forecasting benchmark. The vendor submitted ci-proto results, merged by a maintainer on Nov 11: 51.51 SQL skill and 40.80 MASE skill zero-shot, with no fev-bench datasets in training. The company chart puts ci-proto's mean SQL near 1.16, below Timestral-209M, t0-gamma and TimesFM-3.0 (lower is better), and plots five sizes from XS to XL, though only the 167M model is public. The "#1" claim is the vendor's; check the fev-bench leaderboard.
Getting started
Run pip install continuousintelligence, load with ci.load_model("ContinuousIntelligence/ci-proto"), then call model.forecast(history, horizon=24) for quantile forecasts; the repo includes a bike-rental example and a notebook. Pin a commit SHA in production. Note Section 10 of the licence: external products or redistribution need public attribution and a one-time notice. Founder Ben Walker's launch post had about 270K views at writing time, but the GitHub repo has only about 40 stars and the model around 300 downloads, so it is still early.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.