# Eldric Nexus F1 A from-scratch pretrained zero-shot time-series forecaster. 82.50 M parameters, TiRex-2 architecture ([arXiv:2607.01204](https://arxiv.org/abs/2607.01204)): 12 blocks alternating mLSTM and sLSTM, bidirectional over the context, 9 output quantiles on 32-step patches. The weights are our own — pretrained from random initialisation on our corpus, not fine-tuned from anyone's release. The architecture implementation is NX-AI's, redistributed under Apache-2.0 (see `NOTICE`). ## Quickstart — three commands ```bash pip install https://repo.eldric.ai/models/eldric_nexus_f1-1.0.0-py3-none-any.whl eldric-nexus-f1 check --download # fetch + verify the 315 MB weights, once eldric-nexus-f1 forecast --input your.csv --column value --horizon 48 ``` The third command prints a CSV to stdout: one row per forecast step, nine quantile columns `q10 … q90`. `--out file.csv` writes it to a file instead. `--input` takes a CSV (use `--column` to pick the column) or a file with one number per line; `-` reads stdin. For the Python API, the complete runnable example is [`example_forecast.py`](https://repo.eldric.ai/models/example_forecast.py): ```bash curl -fLO https://repo.eldric.ai/models/example_forecast.py python example_forecast.py ``` which prints, on any machine with the package installed: ``` history 512 steps -> forecast 96 steps, shape (9, 96) step q10 q50 q90 truth 0 131.737 136.037 142.067 133.749 1 129.981 134.516 140.418 134.003 95 130.374 136.663 143.566 140.213 MAE of the median 2.186 MAE of repeat-last-value 10.423 (0.21x — under 1.00 is better) truth inside the q10..q90 band: 100.0 % ``` ## Install ```bash pip install https://repo.eldric.ai/models/eldric_nexus_f1-1.0.0-py3-none-any.whl ``` Pure PyPI, no compiler and **no CUDA toolchain** required: every dependency (`torch`, `numpy`, `einops`, `flashrnn`, `xlstm`, `PyYAML`) ships a pure-Python or prebuilt wheel. On CPU and Apple MPS the recurrent kernels run in plain PyTorch; CUDA is used only if you ask for `device="cuda"`. Python 3.11–3.13. ## Check it works ```bash eldric-nexus-f1 check --download ``` This fetches the weights (315 MB) to `~/.cache/eldric-nexus-f1/`, verifies their SHA-256, loads the model on CPU and prints the parameter count. Run it once before an evaluation sweep — then you know the download and the checksum are fine rather than discovering it mid-run. ## Use it The API is deliberately **identical to TiRex-2's**, so existing evaluation code needs one line changed: ```python from eldric_nexus_f1 import load_model # instead of: from tirex2 import load_model ``` Everything after that is the same: ```python import torch from eldric_nexus_f1 import load_model, TimeseriesType model = load_model() # device="cpu" by default series = TimeseriesType( target=torch.randn(1, 512), # [variates, context], float32 past_covariates=None, future_covariates=None, ) # Returns a LIST, one entry per input series. forecasts = model.forecast([series], prediction_length=64, output_type="numpy") q = forecasts[0][0] # (9, 64) = quantiles x horizon print(q.shape) # (9, 64) print("median of the first 5 steps:", q[4, :5]) print("80 % band at step 0:", q[0, 0], "..", q[8, 0]) ``` Quantile rows are in ascending order, `q[0]` = 0.1 … `q[8]` = 0.9, so `q[4]` is the median. There is no separate mean. **Feed float32.** A float64 target — which is what you get straight out of pandas or `numpy.random` — fails deep inside PyTorch with `mat1 and mat2 must have the same dtype, but got Double and Float`. Cast first: `torch.from_numpy(arr.astype("float32"))`. The CLI does this for you. The return value is a **list**, one entry per input series, each shaped `(variates, quantiles, horizon)`. There is no separate mean — `q50` (index 4) is the median. `forecast_gluon(...)` and `forecast_fev(...)` are available with the same signatures (install the `gluonts` or `fev` extra for those: `pip install "eldric-nexus-f1[fev]"`). From the shell, for a quick look: ```bash eldric-nexus-f1 forecast --input series.csv --horizon 64 --out forecast.csv eldric-nexus-f1 info ``` ## Offline / air-gapped machines Download the checkpoint once and point at it — no network needed afterwards: ```bash curl -fLO https://repo.eldric.ai/models/eldric-nexus-f1-model.ckpt export ELDRIC_NEXUS_F1_WEIGHTS=$PWD/eldric-nexus-f1-model.ckpt ``` `ELDRIC_NEXUS_F1_WEIGHTS` also accepts a directory holding `model.ckpt` plus `model-config.yaml`, which is the layout TiRex-2 itself ships in. The architecture file is included in the wheel, so the checkpoint alone is enough. ## What the model was trained on | | | |---|---| | Parameters | 82,500,000 | | Training steps | 700,000 | | Context length in training | 2048 | | Horizon in training | 320 | | Patch size | 32 (input and output) | | Quantiles | 0.1 … 0.9 in steps of 0.1 | | Real-data corpus | 5.99 M series from LOTSA + Chronos + GIFT-Eval (167.6 GB) | | Synthetic corpus | 15 M Gaussian-process series, drawn at 31.6 % of training samples | | Coupling | the TiRex-2 pretraining coupling chain (paper §3.4 and appendix F): identity, univariate, mixture, linear and non-linear SCM, cointegration, functional | Every figure above is read out of the shipped checkpoint's own `cfg`, not from a training plan. Notably this model saw **no chaotic-system trajectories**: an earlier version of this table claimed it did, which was wrong. It generalises to context and horizon lengths other than the training values; 2048/320 is what the sampler drew during pretraining, not a limit of the model. ## Reproducing the claim that the bundled code is unmodified ```bash pip install "tirex2 @ git+https://github.com/NX-AI/tirex-2@v0.2.1" python pruef/pruef_bitgleich.py ``` Runs the same checkpoint through the bundled copy and through the upstream package on the same input and requires the outputs to be bit-identical. ## Licence Wrapper code and the bundled NX-AI implementation: Apache-2.0 (`LICENSE`, `NOTICE`). The **trained weights** are made available for benchmarking, evaluation and research. For production or commercial use, contact license@eldric.ai. ## Contact - - license@eldric.ai