TiRex-2 Slays Transformers in Time Series: xLSTM Wins

On July 1, 2026, a team of researchers posted TiRex-2 on arXiv, a recurrent xLSTM-based time series foundation model that extends the univariate TiRex to multivariate forecasting with both past and future covariates. This is not just an incremental update—it directly challenges the Transformer-based status quo by handling streaming data without quadratic complexity or full-history retraining.

TiRex-2 is a recurrent xLSTM-based time series foundation model that generalizes univariate TiRex to multivariate forecasting with past and future covariates.
It addresses key limitations of Transformer-based models: quadratic complexity in context length and inability to handle streaming data without full-history retraining.
The model's linear complexity and native streaming support position it as a more practical choice for real-world forecasting where observations arrive continuously.

Why Does TiRex-2 Threaten Transformer Dominance in Time Series?

According to the TiRex-2 paper posted on arXiv on July 1, 2026, existing Transformer-based time series foundation models like TimesNet and PatchTST capture cross-variate dependencies but incur quadratic complexity in context length and require full-history retraining for new data. TiRex-2, built on the xLSTM architecture introduced by Beck et al. in 2024, uses a recurrent design that processes observations sequentially with linear complexity. This means it can handle streaming data—where new observations arrive continuously—without recomputing over the entire history. The paper states that TiRex-2 matches or exceeds Transformer performance on standard multivariate forecasting benchmarks while being significantly more efficient.

What Makes xLSTM a Better Fit for Streaming Forecasting?

TiRex-2 Slays Transformers in Time Series: xLSTM Wins

The xLSTM architecture, as described in the original 2024 paper by Beck et al., modifies the traditional LSTM with exponential gating and a revised memory structure to overcome the vanishing gradient problem. TiRex-2 leverages these improvements to maintain long-range dependencies in multivariate time series without the attention mechanism's quadratic cost. The paper reports that TiRex-2 achieves comparable or better accuracy than Transformer-based models on datasets like ETT and Weather, while requiring 50-70% less computational resources for inference. This makes it particularly suited for edge deployments or real-time systems where latency and memory are constrained.

How Does TiRex-2 Handle Past and Future Covariates?

A critical feature of TiRex-2 is its ability to incorporate both past covariates (known from history) and future covariates (known ahead of time, like holidays or weather forecasts). According to the paper, the model uses a dedicated embedding layer for each covariate type and fuses them with the main time series representation through a learned gating mechanism. This design allows the model to exploit all available information without the need for separate encoders or handcrafted feature engineering. The paper demonstrates that including future covariates improves forecasting accuracy by up to 15% on datasets with known future events, such as the Electricity dataset where holiday schedules are known in advance.

Comparison: TiRex-2 vs. Transformer-Based Models

Feature	TiRex-2 (xLSTM)	Transformer Models (e.g., TimesNet, PatchTST)
Architecture	Recurrent (xLSTM)	Transformer (attention-based)
Complexity in context length	Linear O(L)	Quadratic O(L²)
Streaming data support	Native (sequential processing)	Requires full-history retraining
Multivariate forecasting	Yes, with past and future covariates	Yes, but typically with separate encoders
Computational cost (inference)	50-70% lower	Baseline
Verdict	Winner for streaming and resource-constrained settings	Winner for batch processing with long sequences

What Are the Practical Implications for Forecasting Pipelines?

For organizations deploying time series models in production, TiRex-2 offers a compelling alternative to Transformer-based models. Its linear complexity means it can scale to longer histories without exponential cost increases, and its native streaming support eliminates the need for batch retraining when new data arrives. The paper reports that TiRex-2 achieves state-of-the-art results on the M4 and M5 forecasting competitions, which include millions of time series. This suggests that the model is not just a theoretical improvement but a practical tool for large-scale forecasting tasks.

My Analysis: TiRex-2 is the first serious challenger to the Transformer's hegemony in time series foundation models. While Transformers excel at capturing global dependencies in fixed-length sequences, they are fundamentally ill-suited for the streaming, continuous nature of real-world forecasting. The xLSTM architecture, with its linear complexity and gated memory, is a more natural fit. I believe that within the next 12 months, we will see at least one major cloud provider (AWS, GCP, or Azure) integrate TiRex-2 or a similar xLSTM-based model into their managed time series forecasting services. The winners will be organizations with real-time forecasting needs and limited compute budgets; the losers will be purely Transformer-based model providers that cannot adapt to streaming workloads. A key uncertainty remains: how TiRex-2 handles extremely long-range dependencies (e.g., yearly seasonality) compared to Transformers with learned positional encodings. The paper does not provide ablation studies on this, so I am cautious about claiming a universal win.

Prediction 1: By Q2 2027, AWS will announce a managed time series forecasting service based on an xLSTM architecture, citing TiRex-2's streaming capabilities and cost efficiency as key differentiators.
Prediction 2: The research team behind TiRex-2 will release a closed-source, production-optimized version within 6 months, targeting financial and energy sector clients with high-frequency data.
Prediction 3: At least one major open-source time series library (e.g., Darts, Kats) will add native support for TiRex-2 by the end of 2026.

May 2024
xLSTM architecture introduced
Beck et al. publish the xLSTM paper, modifying traditional LSTMs with exponential gating.
July 2026
TiRex-2 paper released on arXiv
The TiRex-2 model is posted, extending univariate TiRex to multivariate forecasting with past and future covariates.

Insight 1: TiRex-2's linear complexity is not just a theoretical advantage—it translates to real cost savings in production, especially for long-context forecasting tasks.
Insight 2: The model's ability to handle future covariates natively reduces the need for complex feature engineering pipelines, lowering deployment friction.
Insight 3: The streaming support means TiRex-2 is better aligned with the operational reality of most forecasting systems, which receive data in real-time.
Insight 4: The paper's lack of ablation on very long-range dependencies leaves a potential vulnerability for Transformers to exploit in niche applications like climate modeling.
Insight 5: The competition between recurrent and Transformer architectures in time series is now a two-horse race, with implications for the broader foundation model ecosystem.