This repository provides installation and usage scripts for TimeOmni-1. >

🙋 Please let us know if you find out a mistake or have any suggestions!

>

🌟 If you find this resource helpful, please consider to star this repository and cite our research:

🎬 Demo

▶️ Click the teaser to watch the full 2.5-minute demo. It shows all four TSR-Suite tasks (scenario understanding, causality discovery, event-aware forecasting, decision making) answered live by TimeOmni-1-7B on the Hugging Face Space.

---

Updates/News:

🚩 News (Jun. 2026): We have added a demo video showing how TimeOmni-1 solves time series reasoning tasks. 🚩 News (Apr. 2026): We have released new post-trained versions based on Qwen3.5 on Hugging Face: TimeOmni-1-9B and TimeOmni-1-4B. These new versions further scale up model performance, and inference & evaluation are now fully supported by this repository (see below). 🚩 News (Feb. 2026): Please find the open source model on Hugging Face: TimeOmni-1-7B; see also our online demo: https://huggingface.co/spaces/anton-hugging/TimeOmni-1 🚩 News (Jan. 2026): TimeOmni-1 has been accepted to ICLR 2026! 🎉

📊 Benchmarks

Table. Model Size Scaling Comparison

• Note: All metrics below are computed only on valid responses. “–” indicates a success rate (SR) below 10%; in such cases, results are omitted due to insufficient statistical significance, and we therefore do not report them. For ACC, higher is better; for MAE, lower is better. Bold marks the best value in each ACC/MAE column.

	Task1 ID (ACC↑/SR)	Task1 OOD (ACC↑/SR)	Task2 ID (ACC↑/SR)	Task2 OOD (ACC↑/SR)	Task3 ID (MAE↓/SR)	Task3 OOD (MAE↓/SR)	Task4 ID (ACC↑/SR)	Task4 OOD (ACC↑/SR)
7B (Qwen2.5-Instruct)
Qwen2.5-Instruct-7B	48.5/100.0	42.8/100.0	21.6/99.8	26.3/100.0	23.28/53.1	146.12/55.5	25.5/100.0	24.9/100.0
TimeOmni-1-7B	90.7/97.5	87.7/98.3	69.3/99.8	64.0/99.8	14.30/93.8	145.53/82.3	47.9/100.0	58.9/100.0
4B (Qwen3.5)
Qwen-3.5-4B	0.0/16.5	5.9/17.0	28.3/12.4	35.4/12.0	-/2.2	-/9.0	-/8.5	-/9.2
TimeOmni-1-4B	91.5/99.5	91.2/98.4	71.1/100.0	66.1/99.9	13.68/97.6	170.41/86.1	58.5/100.0	72.0/100.0
9B (Qwen3.5)
Qwen-3.5-9B	91.2/51.0	93.5/46.1	43.3/12.1	36.3/12.8	17.56/14.1	-/0.8	64.2/28.2	72.0/32.2
TimeOmni-1-9B	93.5/100.0	92.8/99.8	70.9/100.0	66.2/100.0	13.54/97.8	140.06/95.6	59.6/100.0	75.6/99.6

🛠️ Installation

conda create -n timeomni python=3.10
 conda activate timeomni
 pip install -r requirements.txt

📦 Model Download

Download a single checkpoint by its short tag (7B, 4B, or 9B):

python install/downloadhfmodel.py --model 7B   # anton-hugging/TimeOmni-1-7B (Qwen2.5)
 python install/downloadhfmodel.py --model 4B   # TimeOmni-1/TimeOmni-1-4B    (Qwen3.5)
 python install/downloadhfmodel.py --model 9B   # TimeOmni-1/TimeOmni-1-9B    (Qwen3.5)

Or download all released checkpoints at once:

python install/downloadhfmodel.py --all

Models are stored in the shared Hugging Face cache (.hf/ at the workspace root, i.e. --cache_dir). A --model value may also be any full Hugging Face repo id. To additionally materialise a flat copy, pass --local_dir <path>.

🧪 Dataset Download

python install/download_testbed.py

This creates:

• data/timeomni1idtest.json

• data/timeomni1oodtest.json

🚀 Inference (single question)

Default system prompt:

Output Format:
 <think>Your step-by-step reasoning process that justifies your answer</think>
 <answer>Your final answer(Note: Only output a single uppercase letter of the correct option)</answer>

Run (works for all checkpoints — TimeOmni-1-7B, TimeOmni-1-4B, TimeOmni-1-9B):

python inference/inference.py \
   --model_dir "Local Model Path /models--anton-hugging--TimeOmni-1-7B/snapshots/<hash>" \
   --question "Your Question" \
   --system_prompt "Output Format:\n<think>Your step-by-step reasoning process that justifies your answer</think>\n<answer>Your final answer(Note: Only output a single uppercase letter of the correct option)</answer>"

--model_dir accepts either a local snapshot path or a Hugging Face repo id (e.g. TimeOmni-1/TimeOmni-1-4B). The script auto-selects the right loader: the Qwen3.5-based 4B/9B checkpoints are VLM-backed, and only their text (language-model) weights are loaded for time-series reasoning.

Note (Qwen3.5 / 4B & 9B): these models use linear (gated-delta-net) attention, which

vLLM JIT-compiles at load time and therefore requires the ninja build tool (installed via

requirements.txt; the binary must be on PATH). The evaluation script automatically loads

only the language model for these VLM-backed checkpoints.

📊 Evaluation

bash eval/run-timeomini_test.sh

Optional env overrides:

MODEL_DIR=anton-hugging/TimeOmni-1-7B \
 ANSIDPATH=answer/timeomni1test/yourid_outputs.json \
 RESIDPATH=answer/timeomni1test/yourid_results.json \
 ANSOODPATH=answer/timeomni1test/yourood_outputs.json \
 RESOODPATH=answer/timeomni1test/yourood_results.json \
 bash eval/run-timeomini_test.sh

We report Success Rate (SR), defined as the proportion of model outputs that yield a valid and extractable answer. All other metrics are computed on valid cases only.

• Tasks 1, 2, 4: model outputs a single uppercase letter (A/B/C/D). Metric: Accuracy (ACC).

• Task 3: model outputs a sequence (e.g., [2, 20, 21, ..., 83]). Metric: Mean Absolute Error (MAE).

✍️ Citation

@inproceedings{
 guan2026timeomni,
 title={TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models},
 author={Tong Guan and Zijie Meng and Dianqi Li and Shiyu Wang and Chao-Han Huck Yang and Qingsong Wen and Zuozhu Liu and Sabato Marco Siniscalchi and Ming Jin and Shirui Pan},
 booktitle={The Fourteenth International Conference on Learning Representations},
 year={2026},
 url={https://openreview.net/forum?id=kOIclg7muL}
 }