Chronos-Bolt
The production Chronos: patch-based direct forecasting, 250x faster than the original.
About
Chronos-Bolt (Amazon, November 2024) is what Chronos looks like when it is engineered for production. The original Chronos tokenizes each time step and autoregressively samples trajectories, which is accurate but slow. Bolt chunks the context into patches and decodes quantile forecasts for all future steps at once (direct multi-step forecasting), which Amazon reports as 5% lower error, up to 250 times faster inference, and 20 times better memory efficiency than the original Chronos at the same size. Four sizes ship on Hugging Face from 9M (tiny) to 205M (base) parameters.
Bolt is univariate-only with a 2,048-step context and 64-step prediction length, and it slots into the same amazon-science/chronos-forecasting package and SageMaker JumpStart deployment path as the rest of the family. The honest caveat comes from Amazon's own later release: Chronos-2 beats Bolt head-to-head over 90% of the time and adds multivariate and covariate support, so Bolt is now the fast cheap option rather than the flagship.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported (Amazon, Nov 2024): 5% lower error, up to 250x faster, 20x more memory-efficient than original Chronos at the same size. Amazon's Chronos-2 report notes a >90% win rate for Chronos-2 over Bolt head-to-head.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: amazon-science/chronos-forecasting repo (6,111 stars, Apache 2.0), Hugging Face model cards (amazon/chronos-bolt-tiny through base), provider-published benchmarks, SageMaker JumpStart availability. Strengths: genuinely large efficiency win over the original Chronos, clean production deployment story, permissive licensing. Docked for: (1) efficiency and accuracy numbers are provider-reported; (2) univariate only, no multivariate or native covariate support; (3) superseded as the flagship by Chronos-2. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Amazon's production Chronos (Nov 2024): patch-based direct multi-step forecasting; provider-reported 5% lower error, up to 250x faster, 20x more memory-efficient than original Chronos.
- · HF checkpoints 9M-205M; SageMaker JumpStart deployment; Apache 2.0.
- · Univariate only with no native covariate support; efficiency claims are provider-reported; superseded as flagship by Chronos-2.