Mistral发布Timestral时间序列基础模型家族

Mistral推出了Timestral，其首个时间序列基础模型家族，包含参数规模从5M至209M的四个版本1。最大的209M参数模型在GIFT-Eval、TIME和fev-bench三项主要公开基准测试中均排名第一，超越现有最优方案的同时，参数数量仅为其三分之二1。这些模型支持多变量概率预测、分类协变量和已知未来协变量的处理1。

在性能表现上，28M模型在GIFT-Eval和TIME基准上击败了约四倍规模的Chronos-2模型，而209M模型则以更少参数量超越TimesFM-3和Falcon-2.01。在标准服务器CPU上，5M模型的预测延迟为10毫秒，209M模型为140毫秒，展现出高效的计算能力1。用户可通过GitHub仓库在自有数据上运行Timestral，同时也可利用Mistral Forge平台进行模型微调以实现特定领域的适配1。

Mistral has unveiled Timestral, its inaugural family of time-series foundation models designed to advance forecasting capabilities across various scales.1 The Timestral lineup comprises four models ranging from 5 million to 209 million parameters.1

The largest model in the family, featuring 209 million parameters, achieved top rankings across three major public benchmarks: GIFT-Eval, TIME, and fev-bench.1 This performance milestone is particularly notable given that the model operates with only two-thirds the parameters of the previous state-of-the-art solutions.1 The 28-million-parameter variant outperformed Chronos-2, a model approximately four times larger, on both GIFT-Eval and TIME benchmarks.1 Meanwhile, the 209-million-parameter model surpassed both TimesFM-3 and Falcon-2.0 despite requiring fewer parameters.1

The models support multivariate probabilistic forecasting and can incorporate past and known future covariates into predictions.1 Latency performance on standard server CPUs demonstrates practical efficiency, with the 5-million-parameter model delivering predictions in 10 milliseconds and the 209-million-parameter model in 140 milliseconds.1 Users can deploy Timestral through a publicly available GitHub repository and fine-tune models for domain-specific applications via Mistral Forge.1