Google Releases TimesFM 3.0, a 0.3B-Parameter Time-Series Forecasting Model, in PyTorch
Google Research has published PyTorch weights for TimesFM 3.0, a 0.3-billion-parameter foundation model built for time-series forecasting. The model uses a Stacked Mixing Transformer architecture and is released under a non-commercial license.
Google Publishes TimesFM 3.0 Weights
Google Research has released the official PyTorch weights and configuration files for TimesFM 3.0, the latest version of its time-series foundation model, on Hugging Face. The model is designed specifically for time-series forecasting rather than general text or multimodal tasks.
Architecture and Specifications
TimesFM 3.0 uses a Stacked Mixing Transformer architecture with Variate Attention and CPM Iterative RevIN. According to the model card, key specifications include:
- Parameters: 0.3B
- Layers: 20 transformer layers
- Model dimension: 1280
- Attention heads: 16
- Context patch length: 32
- Forecast horizon patch length: 64
- Output format: Quantile forecasts, with the median at index 4
- Tensor type: F32, distributed as safetensors
This architecture differs from the decoder-only design described in the original TimesFM paper (arXiv:2310.10688, published October 2023), suggesting a structural revision alongside the version bump.
Training Data
Google states that TimesFM 3.0 was pretrained on GiftEvalPretrain, with datasets that overlap with the fev-bench benchmark excluded to avoid contamination. Additional training data includes:
- Wikipedia Pageviews data, with a cutoff of November 2023
- Google Trends top queries data, cutoff end of year 2022
- Synthetic and augmented time-series data
Google has not disclosed the total size of the training corpus or the exact composition breakdown between these sources in the released model card.
Licensing and Availability
The model is released under the TimesFM Non-Commercial License v1.0, meaning it cannot be used in commercial products or services without separate licensing arrangements from Google. As of publication, no inference provider has deployed the model for hosted access — Hugging Face's model page shows it is not currently served through its Inference Providers marketplace, and pricing has not been disclosed since the model is not commercially licensed.
No standardized benchmark scores (such as MASE or CRPS on common forecasting benchmarks) were included in the released model card, so independent verification of forecasting accuracy against competing models like Amazon's Chronos or Nixtla's TimeGPT is not yet available from this source.
What This Means
TimesFM 3.0 extends Google's push into foundation models for structured, non-language data — a category that has received far less attention than LLMs but has real demand in finance, supply chain, energy, and IT operations forecasting. The non-commercial license limits immediate enterprise adoption, positioning this release primarily for researchers and evaluators rather than production deployment. The absence of published benchmark numbers means claims about forecasting quality remain unverified until independent testing against established time-series benchmarks like GIFT-Eval or fev-bench is conducted.
Related Articles
DeepSeek Releases V4-Flash-Vision-Exp, First Multimodal Model in V4 Family
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family, adding visual understanding to the V4-Flash architecture. The 305B-parameter model shows substantial gains on multimodal agent benchmarks while holding steady on text-only tasks.
Zhipu AI Releases GLM-5.3-Flash: First Multimodal Model in GLM-5 Series, 320B Parameters with Only 18B Active
Zhipu AI has released GLM-5.3-Flash, the first natively multimodal model in its GLM-5 series, built on a 320B-parameter mixture-of-experts architecture with only 18B active parameters. The company claims it outperforms GLM-5.2 across benchmarks at one-tenth the cost while approaching Claude Opus 4.8 on coding and agentic tasks.
Tencent Releases WeMM-Embedding-9B, a Multimodal Embedding Model Built on Qwen3.5
Tencent has released WeMM-Embedding-9B, a 9-billion-parameter multimodal embedding model built on Qwen3.5 that produces 4,096-dimensional embeddings from text, images, video, and visual documents. The model reports state-of-the-art results on the MMEB-v2 and MMEB-v3 benchmarks and is released under Apache 2.0.
IBM Releases Granite 4.2 8B, a Dense Reasoning Model with 131K Context and Three Thinking Modes
IBM has released Granite 4.2 8B, a dense reasoning model built for math, code generation, and agentic workflows. The model supports 131K context, 12 languages, and three switchable reasoning modes, priced at $0.10 per 1M input tokens and $0.15 per 1M output tokens.
Comments
Loading...