TimeBraid: Unifying Time Series and Language for Understanding and Forecasting
Abstract
We present TimeBraid, a series of unified time-series and language models that align pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. Each model inherits knowledge, instruction following, and reasoning from one side, continuous-signal perception and zero-shot forecasting from the other, and fuses the two in a shared representation space where both modalities are understood and generated. We study the design choices that make such unified modeling work: where to align the two representation spaces, how to ground language in temporal structure, how to balance understanding with generation, and how to keep joint optimization stable. The resulting recipe combines a unified prompting scheme for diverse time-series and text tasks, stabilized joint training, and supervision from 2.2M curated series--text pairs and 4.9M instruction-tuning samples. Across benchmarks spanning time-series perception, understanding, reasoning, and both context-aided and unimodal forecasting, TimeBraid remains competitive with far larger general-purpose models and task-specific counterparts.
Community
TimeBraid is out: a family of unified time-series + language models
The idea: language models know the world but not the signal; time-series foundation models know the signal but have no way to take in text. TimeBraid keeps both, a pretrained Qwen3 LLM and a TimesFM-based time-series model, and weaves them together layer by layer with shared causal attention. Series stay as continuous patches, text stays as tokens, and both live on one timeline.
So you can hand it a signal and ask about it, or tell it what's coming and watch the forecast move with the text.
Highlights (12 benchmarks):
→ Time-MMD contextual forecasting: #1 average rank of 20 models across 9 domains
→ CGTSF and CAF: best macro MSE/MAE, and the lowest CRPS (GPT-5.4 included on CAF)
→ TemporalBench: TimeBraid-6.7B 36.6% vs 38.5% for GPT-5.4, on par, and ahead of Claude Sonnet 4, Gemini 2.5 Flash and GPT-4o
→ TimeSeriesExam: 50.1 → 60.1 → 63.1 as we scale 1.2B → 2.5B → 6.7B
→ Ctrl-F: the only non-LLM model whose forecasts follow the text condition
TimeBraid-2.5B weights + inference code are open (Apache-2.0):
🌐 Project page & demos: https://xinyuewangg.com/projects/timebraid/
💻 Code: https://github.com/CharonWangg/TimeBraid
🤗 Model: https://huggingface.co/XinyueWangg/TimeBraid-2.5B
📄 Paper: https://arxiv.org/abs/2609.29792
We'd love to hear your feedback and are happy to answer any questions!
Get this paper in your agent:
hf papers read 2609.29792 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 2
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper