Mistral AI releases Timestral time-series forecasting models

Mistral AI's Timestral family ranges from 5 million to 209 million parameters, and the flagship 209M model claims the top spot on three independent time-series benchmarks among pretrained, zero-shot models: GIFT-Eval, TIME, and fev-bench ^[2][3]. On GIFT-Eval, Timestral-209M posts a 0.658 MASE score, edging out Google's TimesFM-3 at 0.667 and beating Amazon's Chronos-2 at 0.698; on the TIME benchmark the gap is similar, 0.626 versus 0.640 and 0.662 ^[2]. What makes the numbers notable is where they run: entirely on a standard CPU, no GPU required. On an 8-thread server CPU, the 5M model produces a one-day forecast from three weeks of 15-minute data in about 10 milliseconds, and the 209M model does it in about 140 milliseconds ^[1]. The models accept up to 32,768 historical time steps as context and can forecast up to 1,024 steps in a single pass, with probabilistic output spread across 21 quantiles, and they support multivariate series plus past-only and known-future covariates ^[2][3]. Perhaps more striking than the flagship's win is that the 28M model - about one-seventh the size of the 209M variant - already outperforms the much larger Chronos-2 on both GIFT-Eval and TIME, and the mid-sized 123M model comes within 0.2% of the 209M model's accuracy while running at about 75% of its latency ^[1]. For enterprises running forecasting at the edge or on commodity infrastructure, that is the real headline: state-of-the-art accuracy without a GPU bill.

Timestral's benchmark rival is not just Google or Amazon - it is a smaller, Austria-based company with an unusually direct line to Mistral's own team. NXAI, the company behind the xLSTM-based TiRex-2 forecasting model, has pushed back on Timestral's architecture, arguing that TiRex-2's streaming design does not need to recompute the entire history at each step, so its cost per step stays constant, unlike Timestral's patch-based transformer approach ^[4]. That critique carries extra weight because of the companies' shared history: Emmi AI, a time-series AI spin-off of NXAI founded at the end of 2024, was later acquired by Mistral, creating a direct lineage between the team that built TiRex and the team now shipping Timestral ^[4]. On the published numbers, Timestral still wins on accuracy - on fev-bench its skill score of 50.4 beats TiRex-2's 45.5, placing TiRex-2 seventh on that leaderboard, and on GIFT-Eval Timestral's 0.658 MASE beats TiRex-2's 0.697 ^[4]. But NXAI's efficiency argument is about a different axis than Mistral's accuracy claims, and both companies are chasing the same industrial customer base, which suggests the architecture debate between patch-based transformers and streaming xLSTM models is far from settled.

Timestral's code is released under the permissive Apache License 2.0, but the model weights - distributed separately on Hugging Face - carry a Modified MIT license with a corporate revenue ceiling ^[1][3]. Once a company's global consolidated monthly revenue exceeds $20 million for the preceding month, it loses the right to use Timestral or any fine-tuned derivative of it, and must instead negotiate a commercial license with Mistral or route through Mistral's own hosted services ^[5]. For smaller companies and researchers the license is effectively open; for any enterprise crossing that revenue line, Timestral quietly becomes a commercial product gated behind Mistral's discretion, a detail easy to miss amid the headline CPU-only, beats-TimesFM-3 framing.

Not every reaction has been enthusiastic. On a technical forum discussion thread, the top comment dismissed the model as complete snake oil for commodity and price forecasting, arguing that geopolitical shocks simply cannot be captured by historical trend data no matter how sophisticated the model. Other commenters pushed back that a general-purpose world-model approach should generalize better than classical statistical baselines like ARIMA; one user said the license terms alone would make them pick Timestral over TimesFM, and another questioned whether a dedicated forecasting model is worth the overhead versus having an LLM write custom forecasting code on demand. The debate echoes a long-standing tension in time-series forecasting: benchmark wins on backtested public datasets do not necessarily translate into reliability for use cases like commodity pricing, where the biggest moves are driven by events no model has seen before. Mistral's own benchmark claims against GIFT-Eval, TIME, and fev-bench are self-reported, with no independent third-party assessment available at the time of coverage ^[6], which adds another reason for the skepticism to persist until outside users publish their own results.