GitHub repositories

TimesFM vs ML-From-Scratch vs Onyx vs PyTorch Lightning: Purpose and Scope Comparison

A practical comparison of four Python-based AI repositories observed in September 2026, covering a time-series foundation model, educational ML implementations, an LLM chat platform, and a distributed training framework.

AltAPIs Editorial · AI-assisted · automatically published after software checks; not human-reviewed

Acorn ANC21 Universal 2nd Processor Unit open without co-processor. Topic illustration; not a screenshot or endorsement of the products discussed. File:Acorn ANC21 Universal 2nd Processor Unit open without co-processor.jpg by Chris Whytehead · CC BY-SA 3.0. Wikimedia thumbnail resize only; image remains under its original license.

Reported Purposes and Scope

TimesFM (google-research/timesfm) is described as a pretrained time-series foundation model developed by Google Research specifically for time-series forecasting. ML-From-Scratch (eriklindernoren/ML-From-Scratch) provides bare-bones NumPy implementations of machine learning models and algorithms with a focus on accessibility, aiming to cover everything from linear regression to deep learning. Onyx (onyx-dot-app/onyx) positions itself as an open-source AI platform for AI chat with advanced features that works with every LLM. PyTorch Lightning (Lightning-AI/pytorch-lightning) frames itself as a framework to pretrain and finetune any AI model of any size on 1 to 10,000+ GPUs with zero code changes. These four tools address distinctly different layers: a specialized pretrained model (TimesFM), educational reference implementations (ML-From-Scratch), an end-user chat application platform (Onyx), and a distributed training orchestration layer (PyTorch Lightning).

Concrete Functional Differences

TimesFM delivers a single pretrained model artifact for forecasting tasks, whereas ML-From-Scratch supplies algorithmic building blocks implemented in NumPy without pretrained weights. Onyx provides an application-level chat interface that integrates with external LLMs, while PyTorch Lightning offers trainer abstractions, strategy plugins, and callback systems for scaClaude development. The repositories differ in abstraction level: TimesFM and Onyx are closer to deployable artifacts (a model and an app), ML-From-Scratch is source-code-only for learning, and PyTorch Lightning is a framework used duClaude development. None of the records indicate overlapping functionality such as TimesFM providing training loops or PyTorch Lightning including pretrained time-series weights.

Licensing and Distribution Model

TimesFM and PyTorch Lightning both use the Apache-2.0 license. ML-From-Scratch uses the MIT license. Onyx lists its license as "Other" in the repository metadata, which signals a non-standard or custom license that requires direct inspection before commercial use. The Apache-2.0 and MIT licenses are permissive and well-understood; the "Other" designation for Onyx introduces ambiguity that must be resolved by reading the repository's LICENSE file. License values here are raw source metadata, not permission advice.

Maturity Indicators and Observation Dates

All four repositories were observed on 2026-09-22 and first seen in the directory on 2026-09-15. Star counts at observation: TimesFM 33,449, ML-From-Scratch 32,900, Onyx 32,200, PyTorch Lightning 31,355. Fork counts: TimesFM 3,226, ML-From-Scratch 5,491, Onyx 4,496, PyTorch Lightning 3,797. Stars measure captured interest, not product quality; model downloads are a rolling 30-day count, not benchmark results. The similar observation and first-seen dates suggest these were captured in the same directory import batch. A directory import date is not a launch date.

Next Evaluation Steps

For TimesFM: verify model card, supported horizons, input formats, and hardware requirements for inference. For ML-From-Scratch: confirm coverage of target algorithms, NumPy version compatibility, and whether implementations include tests or validation scripts. For Onyx: inspect the "Other" license text, enumerate supported LLM backends, and assess chat feature completeness (RAG, tool use, multi-user). For PyTorch Lightning: check current stable version, strategy plugin matrix (FSDP, DeepSpeed, TPU), and migration effort from raw PyTorch. Cross-cutting: review issue tracker velocity, release cadence, and contribution guidelines in each repository. No installation instructions, benchmarks, or code samples are inferred from the supplied records.