Mistral Releases Timestral: Four Time-Series Models From 5M to 209M Parameters

Mistral shipped four public forecasting checkpoints

Mistral announced Timestral on October 15, 2026 as its first family of time-series foundation models. At review time, mistralai/Timestral-5M, mistralai/Timestral-28M, mistralai/Timestral-123M and mistralai/Timestral-209M were public and ungated on Hugging Face. Their repository records were created on October 14, while Mistral’s public inference repository was created on October 15.

These are downloadable artifacts, not a promise of future weights. Mistral lists FP32 downloads of 21 MB, 111 MB, 490 MB and 837 MB. The standalone package requires Python 3.10 or newer, PyTorch 2.4 or newer, NumPy, safetensors and huggingface_hub, and it can load a pinned Hub revision or local cached files for offline inference.

The output is a distribution, not one point estimate

Timestral accepts univariate or multivariate target series, missing values, past-only covariates and covariates whose future values are already known. Categorical covariates are also supported. Each forecast returns 21 quantiles from 0.01 through 0.99, including the median, so an application can retain a range rather than pretending one predicted value is certain.

Every checkpoint supports up to 32,768 historical steps. A single forward pass covers up to 1,024 future steps. For a longer horizon, the implementation forecasts block by block and appends each completed block’s median to the context. That recursive route can accumulate error, so “supports a longer horizon” should not be read as verified long-range accuracy.

The architecture separates time, variables and forecast tokens

The 209M model is a 24-layer patch-based transformer with 768-dimensional embeddings and 32 time steps per patch. Mistral calls its training approach Decoupled Horizon Forecasting: separate prediction tokens read the causally encoded history, preserving observed values as context while keeping genuinely missing values distinct.

Covariate-aware attention gives each input row a role. Known-future covariates can expose their complete window, targets stay causal, and information is designed to flow from covariates to targets rather than back into covariates. Mistral says most pre-training data was synthetic, followed by post-training on real and synthetic data with four times the initial context. A promised technical report is not yet public, so the full corpus, ablations, contamination controls and reproducibility detail remain unavailable.

Benchmark leadership is a vendor claim pending reproduction

Mistral reports evaluating Timestral zero-shot on GIFT-Eval, TIME and fev-bench with one checkpoint and no ensemble. The 209M card reports GIFT-Eval MASE of 0.658 and CRPS of 0.446, TIME MASE of 0.626 and CRPS of 0.522, and fev-bench SQL and MASE skill scores of 50.39 and 39.49. Lower is better for the first four values; higher is better for the skill scores.

Those tables compare Timestral with named public models, and Mistral says results were submitted to the GIFT-Eval and TIME leaderboards. AccessAllGPT did not recreate the environments, inspect every task-level result or verify statistical uncertainty. Forecasting performance can move with frequency, horizon, missingness, covariate availability, regime shifts and leakage controls, so the launch averages are a trial hypothesis rather than a deployment guarantee.

CPU inference is plausible, but the published timing is narrow

Mistral says no GPU is required. Its launch article reports that an 8-thread standard server CPU forecasts one day from three weeks of 15-minute history in 10 milliseconds with the 5M model and 140 milliseconds with the 209M model. It also says the 28M version is five times faster than 209M and the 123M version comes within 0.2% of its score at three quarters of the latency.

These are provider measurements for one stated shape, not a hardware-neutral service objective. A production trial should record exact CPU, thread count, batch size, number of variates and covariates, missingness, horizon, process concurrency, cold start, peak memory and p95 latency. The smallest model may be the better system if it meets the workload threshold with materially lower operating cost.

Code and weights have different licenses

Mistral licenses the inference package and tests under Apache 2.0. The model weights are separate: their Modified MIT license requires attribution and says a user is not authorized to exercise the license rights if its company—or employer—had more than $20 million in global consolidated revenue in the preceding month. Mistral offers a commercial-license request route and points to its hosted services.

That revenue condition means “public weights” is more accurate than an unqualified “open source” label for procurement. Teams should preserve both license texts with the pinned code and model commits, identify the legal entity and revenue test that applies, and obtain qualified legal review before redistribution, modification or production use. No self-serve hosted Timestral price, token meter or service level was published in the reviewed launch materials.

What is available now—and what remains missing

Available now are four public checkpoints, standalone inference code, CPU and GPU execution, a forecast API, covariate handling, commit pinning and offline cache use. Mistral also says Timestral can be adapted through Mistral Forge, including fine-tuning on proprietary data, changing the training mix and deploying the pipeline on client hardware or in air-gapped environments.

The public launch does not supply the promised technical report, a managed API identifier, self-serve hosted price, quota, regional matrix or service-level agreement. Forge access and commercial terms are not established by the downloadable repository. Teams should treat public local inference, negotiated adaptation services and any future hosted endpoint as three separate product decisions.

What forecasting teams should do next

Start with the 5M or 28M checkpoint against a simple seasonal or statistical baseline and the current production forecaster. Freeze a rolling-origin evaluation across representative frequencies, horizons, missingness and regime changes. Score point accuracy, quantile calibration, interval coverage, decision cost, latency, memory and operator effort; preserve results by segment rather than one aggregate.

Advance to 123M or 209M only when measured accuracy or calibration justifies the added footprint. Constrain recursive long-horizon use until drift is quantified. Wait for the technical report when training provenance or benchmark reproduction is a mandatory gate. Reject the weights when the license does not fit and no commercial agreement is in place; public downloadability is not deployment authorization.

Disclosures

AccessAllGPT received no weights, Mistral Forge access, hosted access, credits, hardware, data, benchmark logs, private briefing, review, payment or compensation for this article. Mistral AI and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Timestral. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Mistral AI or Hugging Face. Publication-wide relationships are listed on the disclosures page.