Mistral releases Timestral time-series models from 5M to 209M parameters, forecasting a day ahead in 140 ms on a CPU
Mistral’s first time-series foundation models go small enough to skip the GPU.
// Key points
- Four sizes, 5M, 28M, 123M and 209M, share one architecture and recipe; they produce probabilistic multivariate forecasts with 21 quantiles, take past-only and known-future covariates, and have a 32,768-step context.
- The 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench with a single checkpoint and no ensembling; the 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME.
- On an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
- The license is Modified MIT, with a GitHub repo and Hugging Face collection, and it can be used and fine-tuned through Mistral Forge.
Builder's takeCapacity planning and usage forecasting are things we’ve handled with rules or not at all. A zero-shot model that runs in about a hundred milliseconds on a CPU is worth a try: I’ll backtest it on AI Cloud Drive’s storage growth and model call volume from the past few months to see if it can replace our guesswork scaling thresholds. Read what the Modified MIT license changes before going to production.
// Background · from #Open models
Full timeline →- Oct 16 Meta releases open weights for Muse Spark Open, an 816B-total, 42B-active multimodal MoE with a 1M context
- Oct 16 NVIDIA opens the 120B Nemotron 3.5 Super VL with datasets and training recipes; the 74B-A9B Bolt fits in 24GB with native 2-bit quantization
- Oct 15 Microsoft releases UserRepGen-4B under MIT: it turns multi-turn conversations into user descriptions for simulators, scoring 4.95 on fidelity against GPT-5.5’s 3.90