Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Mistral AI Releases Timestral: A 209M Forecasting Model Family - OpenSmartRoute
Mistral AI Releases Timestral: A 209M Forecasting Model Family
Mistral AI launched Timestral, a new family of time series forecasting models ranging from 5M to 209M parameters. The largest model outperforms Google's TimesFM-3 on key benchmarks while running on standard server CPUs.
Key points
Timestral-209M ranks first among zero-shot models on GIFT-Eval, TIME, and fev-bench.
The 209M model achieves a SQL skill score of 50.39, beating TimesFM-3 at 48.99.
Inference runs on standard server CPUs without needing a GPU.
The 5M model forecasts in 10 ms, while the 209M model takes 140 ms.
Why it matters: Engineers can now run high-accuracy forecasting models on cheap CPUs instead of expensive GPUs.
By OpenSmartRoute editorial · written through the router by writer-small
From MarkTechPost - “Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3”
Mistral AI announced a new family of models for forecasting time series data. This release marks their first entry into the time series foundation model space. The project is called Timestral. It offers four different sizes of models. These models range from 5 million to 209 million parameters. They are designed to predict future values based on past data patterns.
Timestral uses a specific type of machine learning architecture. It relies on transformers adapted for sequential data. Unlike older tools, it does not require retraining for every new dataset. This allows it to work immediately on unseen data. The team calls this capability zero-shot performance. Users can deploy these models without extensive preparation.
The models are built to run on standard server hardware. They do not strictly require a graphics processing unit. This lowers the barrier to entry for many organizations. Teams can now forecast data using existing CPU servers. This change removes the need for expensive GPU clusters.
The announcement includes detailed specifications for each model size. Developers can choose the right fit for their needs. Smaller models offer speed, while larger models offer accuracy. The largest model, Timestral-209M, leads in benchmark tests. It beats other top models in key categories.
Title: Mistral AI Unveils Timestral: CPU-Only Forecasting Models Beat Top Competitors
Summary: Mistral AI released Timestral, a family of forecasting models from 5M to 209M parameters. The largest model outperforms Google's TimesFM-3 on benchmarks while running on standard CPUs.
Why: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
Highlights:
Timestral-209M ranks #1 on GIFT-Eval, TIME, and fev-bench among zero-shot models.
Models run on standard server CPUs without needing a GPU.
The 209M model weighs only 837 MB in fp32 format.
Inference speeds range from 10 ms to 140 ms on an 8-thread CPU.
Outline:
What Mistral AI Announced - The release of Timestral as a new foundation model family
Deepseek released V4.1 Flash 1017 which beats larger rivals in coding and agent tasks. The model also cuts per-task costs significantly while producing shorter outputs.
TypeSafe released Jev, a closed decision model, and 20 alternatives appeared within weeks. This guide compares their sizes, licenses and benchmarks.
Model Sizes and Performance - Specific parameter counts, context limits, and benchmark scores
How Timestral Works - Technical architecture including patch-based transformers and attention mechanisms
Comparison with Competitors - Direct performance comparison against TimesFM-3, Chronos-2, and Toto 2.0
Why it Matters - Impact on cost, speed, and accessibility for running forecasting models
How OpenSmartRoute Helps - Configuring the router to deploy Timestral efficiently
What to Do - Steps for developers to install, test, and deploy Timestral
Tags: models, forecasting, inference, cpu, benchmarks
Keywords: Timestral, Mistral AI, time series, foundation model, zero-shot, CPU inference
Mistral AI announced a new family of models for forecasting time series data. This release marks their first entry into the time series foundation model space. The project is called Timestral. It offers four different sizes of models. These models range from 5 million to 209 million parameters. They are designed to predict future values based on past data patterns.
Timestral uses a specific type of machine learning architecture. It relies on transformers adapted for sequential data. Unlike older tools, it does not require retraining for every new dataset. This allows it to work immediately on unseen data. The team calls this capability zero-shot performance. Users can deploy these models without extensive preparation.
The models are built to run on standard server hardware. They do not strictly require a graphics processing unit. This lowers the barrier to entry for many organizations. Teams can now forecast data using existing CPU servers. This change removes the need for expensive GPU clusters.
The announcement includes detailed specifications for each model size. Developers can choose the right fit for their needs. Smaller models offer speed, while larger models offer accuracy. The largest model, Timestral-209M, leads in benchmark tests. It beats other top models in key categories.
Model Sizes and Performance - Specific parameter counts, context limits, and benchmark scores
The Timestral family includes four distinct checkpoints. Each size is dense and optimized for different tasks. The smallest model has 5 million parameters. The next size has 28 million parameters. The third size contains 123 million parameters. The largest model has 209 million parameters.
These models support a context window of 32,768 time steps. This allows them to analyze long historical datasets. A single forward pass can forecast up to 1,024 steps ahead. Longer horizons require running the model in blocks. The system feeds each block's median back as context.
Benchmark results show strong performance across multiple platforms. The 209M model achieved a SQL skill score of 50.39 on fev-bench. This score is higher than Google's TimesFM-3, which scored 48.99. The model also achieved a MASE score of 0.658 on GIFT-Eval. This metric is lower is better, so 0.658 is a strong result.
Performance scales predictably with model size. The 123M model scores 0.659 MASE on GIFT-Eval. This is nearly identical to the 209M model's score of 0.658. The 28M model beats Chronos-2 on GIFT-Eval and TIME. Chronos-2 is approximately four times larger than the 28M model.
Inference speed varies significantly by model size. The 5M model took 10 milliseconds to forecast one day ahead. The 209M model took 140 milliseconds for the same task. The 28M model is five times faster than the 209M model. The 123M model lands within 0.2% of the 209M score. It does this at three quarters of the latency.
Weights are stored in fp32 format for the largest model. The 209M model weighs 837 MB in total. Smaller models require less storage space. The 5M model needs 21 MB for its weights. The 28M model needs 111 MB for its weights. The 123M model needs 490 MB for its weights.
Context limits are consistent across all sizes. Every model accepts up to 32,768 time steps. This covers weeks of 15-minute historical data easily. The models accept past-only, known-future, and categorical covariates. This flexibility supports complex forecasting scenarios.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
How Timestral Works - Technical architecture including patch-based transformers and attention mechanisms
Timestral is a time series foundation model. It is pretrained on a large, diverse corpus of data. The model then forecasts on data it has never seen before. This approach transfers patterns across different domains. Classical tools like ARIMA fit one series at a time. Deep models like PatchTST usually train from scratch on each dataset.
Timestral transfers patterns across domains instead of retraining. Each model outputs 21 quantiles per time step. This provides probabilistic forecasts rather than single point predictions. The system forecasts multivariate targets jointly. It accepts past-only, known-future, and categorical covariates.
The architecture uses a causal, patch-based transformer. It includes time attention and variate attention mechanisms. Each patch covers 32 time steps. The 209M model uses 24 layers. The embedding dimension is 768. This structure allows efficient processing of long sequences.
Mistral lists four technical contributions to the design. The first is Decoupled Horizon Forecasting. Learned forecast tokens read the full history. Observed values stay in context instead of being masked. This differs from contiguous patch masking in Toto 2.0 and TimesFM-3.
The second contribution is covariate-aware attention. Each series acts as a target. It can be a past-only covariate or a known-future covariate. Information flows only from covariates to targets. This prevents future leakage in the predictions.
The third contribution involves synthetic-first data. Most training data is synthetic in nature. It includes trends, seasonality, noise, and nonlinear dependencies. This synthetic data helps the model learn robust patterns.
The fourth contribution is two-stage training. The model pretrains on synthetic data at a context of 8,192. It then post-trains on mixed data with four times longer context. The loss function is pinball loss over 21 quantiles. This ensures accurate probabilistic distributions.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
Comparison with Competitors - Direct performance comparison against TimesFM-3, Chronos-2, and Toto 2.0
Timestral-209M competes directly with several established models. It faces Google's TimesFM-3 and Amazon's Chronos-2. It also faces Toto 2.0 from Meta. Google and Amazon do not publish matching figures on their own model pages. Datadog publishes figures for Chronos-2.
On the GIFT-Eval benchmark, Timestral-209M leads the zero-shot category. Its MASE score is 0.658. TimesFM-3 scores 0.642 according to Mistral's card. TimesFM-3 lists a CRPS score of 0.539. Mistral's card lists a CRPS score of 0.446 for Timestral.
On the TIME benchmark, Timestral-209M also leads. Its MASE score is 0.626. TimesFM-3 scores 0.642 on this benchmark. The CRPS score for TimesFM-3 is 0.522. Timestral's CRPS score is 0.522 as well.
On the fev-bench benchmark, Timestral-209M leads again. Its SQL skill score is 50.39. TimesFM-3 scores 48.99 on this metric. The MASE skill score for Timestral is 39.49. This is higher than TimesFM-3's score.
The size scaling is flat at the top for Timestral. The 123M model scores 0.659 MASE on GIFT-Eval. This is against 0.658 for the 209M model. The difference is negligible. The 28M model beats Chronos-2 on GIFT-Eval and TIME. Chronos-2 is about four times larger.
Toto 2.0 uses contiguous patch masking. This differs from Timestral's decoupled approach. Toto 2.0 is 2.5B parameters in size. Timestral models are much smaller. Timestral offers better accuracy per parameter count.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
Why it Matters - Impact on cost, speed, and accessibility for running forecasting models
Forecasting models have historically required expensive GPU infrastructure. Timestral changes this by running on standard server CPUs. This reduces the total cost of ownership significantly. Teams no longer need to buy dedicated GPU clusters.
Speed is another major advantage. The 5M model runs in 10 milliseconds. The 209M model runs in 140 milliseconds. These times are fast enough for real-time applications. The 28M model is five times faster than the 209M model.
Accessibility is improved by open weights. Developers can download the models easily. They can run the inference code locally. This removes vendor lock-in for many organizations. The models are available on Hugging Face.
Safety is enhanced by modified MIT weights. TimesFM-3 uses a non-commercial licence. Timestral allows broader usage rights. This opens the door for commercial applications.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
How OpenSmartRoute Helps - Configuring the router to deploy Timestral efficiently
Teams using OpenSmartRoute gain immediate value from Timestral. The router sends each request to the best-fit model from a catalogue. It scores every candidate on quality, cost, speed, and safety. The team sets the weights per request.
A new model is simply one catalogue entry. Timestral becomes a new entry in the system. It competes on the next request automatically. Nothing else changes in the app. The router learns from outcomes over time.
The savings ledger shows what each routed request cost. It compares this to what the most expensive model would have cost. This provides clear financial visibility. An input guard spots prompt injection before a request leaves.
OpenSmartRoute works with open-weight models served locally. It supports Timestral running on standard CPUs. The router handles the traffic distribution efficiently. It ensures the right model gets the right request.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
What to Do - Steps for developers to install, test, and deploy Timestral
Developers should start by installing the package. Use the command pip install git+https://github.com/mistralai/timestral.git. This requires Python 3.10 or higher. It also requires torch version 2.4 or higher.
Dependencies include PyTorch, NumPy, safetensors, and huggingface_hub. Weights download from Hugging Face directly. The 5M model is 21 MB. The 28M model is 111 MB. The 123M model is 490 MB.
Test the models with sample time series data. Check the inference speed on your hardware. Verify the accuracy against known benchmarks. Ensure the context window fits your data needs.
Deploy the models in your production environment. You can run them on-premises or in a VPC. The pipeline can also run in an air-gapped environment. Mistral Forge supports fine-tuning on proprietary data if needed.
Check the blog post, GitHub repo, and model weights on Hugging Face. All credit goes to the researcher of this project. Follow the team on Twitter for updates. Join the 150k+ ML SubReddit for community support. Subscribe to the newsletter for more news.
Why it Matters: This news matters because it allows teams to run accurate forecasting models on cheap hardware, saving money and reducing latency.
Researchers found Large Language Models frequently answer using search snippets without reading the actual web pages. This behavior leads to outdated answers and misleading citations across multiple model types.
