Skip to content
CAI
Software that uses CAICheck a score

google-research/timesfm

45.2

Weak · 15 November 2026

17k

lines of production code

Python

with JavaScript

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add MLX (Apple Silicon) backend for TimesFM 3.0

Introduces a new MLX-based inference backend for TimesFM 3.0, enabling native execution on Apple Silicon hardware. This addition includes the core model implementation, configuration mapping, and a forecaster wrapper that mirrors the existing PyTorch API. The backend supports multivariate targets and covariates, iterative CPM RevIN refinement, and various architectural options (e.g., stitching, residual activations). It is designed to be numerically parity-checked against the PyTorch backend and gracefully handles environments where MLX is not available.

src/timesfm3/mlx · high confidence

Add TimesFM 3.0 benchmarking suite and documentation

The timesfm3-usage area now includes a new benchmarks directory containing evaluation notebooks, runners, and results for TimesFM 3.0. This update adds support for the AutoGluon FEV-Bench (100 tasks), the ICML TIME Benchmark (98 tasks), and the Salesforce GIFT-Eval, along with corresponding README documentation and initial result data.

timesfm3-usage · high confidence

Introduce TimesFM 2.5 model implementations in Flax and PyTorch

Added new model implementations for TimesFM 2.5 (200M parameters) in both Flax (\timesfm\_2p5\_flax.py\) and PyTorch (\timesfm\_2p5\_torch.py\). These files provide the core architecture, including tokenizers, stacked transformers, and autoregressive decoding logic, enabling users to load and run the TimesFM 2.5 model using their preferred framework.

_src/timesfm/timesfm\2p5 · high confidence

Behavioural changes

18 commits (6 fixes) modifying src/timesfm3/torch

A change to existing behaviour in src/timesfm3/torch — 18 commits (6 fixs), 16 files: “fix: apply symmetric tolerance guard in revin denormalization”; “Add unit tests for TimesFM3Evaluator”; “Gather CPM anchor patches without one-hot selection”; and 15 more.

src/timesfm3/torch · medium confidence · unverified

33 commits (9 fixes) modifying timesfm

A change to existing behaviour in timesfm — 33 commits (9 fixs), 6 files: “bug fix, docstring, readme”; “fix v1 jax version”; “feat(timesfm3/mlx): support multivariate targets and covariates”; and 30 more (refs \#369).

timesfm · medium confidence · unverified

7 commits (0 fixes) modifying timesfm-forecasting

A change to existing behaviour in timesfm-forecasting — 7 commits, 28 files: “Reject CSV columns without observations before forecasting”; “Keep missing CSV observations on the shared time grid”; “docs: address PR \#369 review comments and add dataset preflight”; and 4 more (refs \#369).

timesfm-forecasting · medium confidence · unverified

7 commits (1 fix) modifying src/timesfm/torch

A change to existing behaviour in src/timesfm/torch — 7 commits (1 fix), 5 files: “fix: apply symmetric tolerance guard in revin denormalization”; “2.0.0 initial”; “Document intentional zero-init of RMSNorm scale (torch)”; and 4 more.

src/timesfm/torch · medium confidence · unverified

7 commits (2 fixes) modifying src/timesfm/flax

A change to existing behaviour in src/timesfm/flax — 7 commits (2 fixs), 5 files: “fix: apply symmetric tolerance guard in revin denormalization”; “(1) Make config hashable. (2) Flax”; “Initialize scale parameter to zeros”; and 4 more.

src/timesfm/flax · medium confidence · unverified

TimesFM API restructured with framework-specific model imports

The public API in src/timesfm/\_\init\\_.py has been restructured to expose framework-specific model classes directly rather than a single auto-detected TimesFm class. Users can now import TimesFM\_2p5\_200M\_torch and TimesFM\_2p5\_200M\_flax from the package, along with TimesFM3Forecaster and TimesFM3Torch from the timesfm3 subpackage, alongside the new ForecastConfig. The previous auto-detection logic that printed warnings and imported a unified TimesFm class based on Python version or availability has been removed.

src/timesfm · high confidence

TimesFM3 PyTorch backend reorganized with backward-compatible imports

The TimesFM3 PyTorch implementation has been moved into a dedicated \\timesfm3.torch\\ subpackage, while the top-level \\timesfm3\\ namespace now lazily re-exports key classes (such as \\TimesFM3Forecaster\\ and \\TimesFM3Torch\\) to ensure existing import paths continue to work without breaking changes. This structural change improves modularity by separating the PyTorch backend from the MLX (Apple Silicon) backend, and includes a comprehensive test suite to verify that legacy import patterns remain fully functional.

src/timesfm3 · high confidence

Fixes

7 commits (6 fixes) fixing src/timesfm/utils

A fix in src/timesfm/utils — 7 commits (6 fixs), 1 file: “fix(xreg): handle empty dimensions in padded arrays”; “covariate support”; “fix: improve PEFT device consistency and XReg output slicing”; and 4 more (refs \#393, \#390).

src/timesfm/utils · medium confidence · unverified

Fixes for v1 data batching, variance calculation, and covariate leakage

This release corrects three bugs in the v1 codebase: the data loader now respects the batch\_size when permute=False, ensuring batches are correctly sized; the variance calculation in the PyTorch decoder is numerically stabilized to prevent negative variance values; and the external regressor (xreg) module now fits ridge regression per-input to prevent data leakage across batch items.

v1 · high confidence

Test coverage

28 commits adding/updating tests in tests

Test coverage in tests — 28 commits (12 fixs), 21 files: “fix(xreg): handle empty dimensions in padded arrays”; “fix: apply symmetric tolerance guard in revin denormalization”; “fix: stop forecast() from mutating the caller's input list”; and 25 more.

tests · high confidence · unverified

Dependencies

Migrate to PyProject Standard and Add MLX Backend for TimesFM 3.0

The main package manifest has migrated from Poetry to the standard pyproject.toml format (using setuptools), bumping the version to 3.0.2 and simplifying core dependencies to numpy, huggingface\_hub, and safetensors. This change introduces a new optional MLX backend for Apple Silicon devices and reorganizes optional dependencies for Torch, Flax, and xreg. Additionally, a new v1 subdirectory now contains a separate Poetry-based configuration for the legacy 1.3.0 version, and a generated requirements.txt is included for compatibility.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

This is the PUBLIC form of this artifact. Findings are listed in full, but the details of SECURITY findings — which rule fired, in which file, on which line, and how to fix it — are deliberately withheld, and any secret-scanner results are excluded entirely. Where detail is absent here it was REMOVED FOR PUBLICATION; it is not missing from the analysis. The complete artifact is available from the repository owner.

Score

  • CAI 44 → 45 (+0.8)
  • Rubric changed (rubric-2026.11.1 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 68 → 61 (-6.4)
  • Architecture 89 → 90 (+1.6)
  • Maturity 81 → 80 (-0.8)
  • Readiness 50 → 52 (+2.2)
  • Security 24 → 27 (+2.4)

Resolved (14)

  • Duplicated block (10–11 lines × 2) (src/timesfm/flax/util.py)
  • Duplicated block (11 lines × 2) (src/timesfm/utils/xreg_lib.py)
  • Duplicated block (11 lines × 2) (src/timesfm3/torch/timesfm3_forecaster.py)
  • Duplicated block (12 lines × 2) (src/timesfm/utils/xreg_lib.py)
  • Duplicated block (12 lines × 2) (src/timesfm/utils/xreg_lib.py)
  • Duplicated block (29–32 lines × 2) (src/timesfm/timesfm_2p5/timesfm_2p5_flax.py)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • High: security finding (details withheld)
  • Off-boarding risk: anonymized user #1
  • Projects may be oversized for their cohesion
  • Release publish has no approval gate
  • Workflow token permissions not restricted (.github/workflows/main.yml)
  • Workflow token permissions not restricted (.github/workflows/manual_publish.yml)

New (47)

  • Duplicated block (10 lines × 2) (src/timesfm/timesfm_2p5/timesfm_2p5_flax.py)
  • Duplicated block (12 lines × 2) (src/timesfm/utils/xreg_lib.py)
  • Duplicated block (14–15 lines × 2) (src/timesfm3/torch/timesfm3_forecaster.py)
  • Duplicated block (15 lines × 2) (src/timesfm3/torch/model.py)
  • Duplicated block (16 lines × 2) (src/timesfm/utils/xreg_lib.py)
  • Duplicated block (27–28 lines × 2) (src/timesfm/timesfm_2p5/timesfm_2p5_flax.py)
  • Duplicated block (6 lines × 2) (src/timesfm3/torch/timesfm3_forecaster.py)
  • Duplicated block (7 lines × 2) (src/timesfm/flax/util.py)
  • Edited copies of one member (3 members, 30+ corresponding lines) (v1/src/adapter/dora_layers.py)
  • Edited copy of a member (11 corresponding lines) (src/timesfm/flax/transformer.py)
  • Edited copy of a member (12 corresponding lines) (src/timesfm/flax/util.py)
  • Edited copy of a member (13 corresponding lines) (src/timesfm/flax/transformer.py)
  • Edited copy of a member (15 corresponding lines) (src/timesfm/timesfm_2p5/timesfm_2p5_flax.py)
  • Edited copy of a member (18 corresponding lines) (src/timesfm/init.py)
  • Edited copy of a member (20 corresponding lines) (src/timesfm/flax/dense.py)
  • Edited copy of a member (20 corresponding lines) (src/timesfm3/mlx/transformer.py)
  • Edited copy of a member (25 corresponding lines) (v1/src/finetuning/finetuning_torch.py)
  • Edited copy of a member (3 corresponding lines) (v1/src/timesfm/patched_decoder.py)
  • Edited copy of a member (31 corresponding lines) (src/timesfm/flax/transformer.py)
  • Edited copy of a member (32 corresponding lines) (src/timesfm/timesfm_2p5/timesfm_2p5_base.py)
  • …and 27 more

Changes since last survey

  • 43 commits — 40 feature/other, 3 fixes

By area

  • (repo) — 37 commits
  • src/timesfm — 2 commits
  • (root) — 1 commit
  • .github/workflows — 1 commit
  • src/timesfm3 — 1 commit
  • timesfm-forecasting/scripts — 1 commit

Notable commits

  • fix: fix(timesfm): keep the optional-backend error an AttributeError
  • fix: fix(timesfm3): report a missing torch backend as an absent attribute
  • fix: fix: surface optional backend import failures with actionable errors
  • change: Merge pull request #365
  • change: Merge pull request #431
  • change: Merge pull request #468
  • change: Merge pull request #478
  • change: Merge pull request #481
  • change: Merge pull request #484
  • change: Merge pull request #486
  • change: Merge pull request #504
  • change: Merge pull request #511
  • change: Merge pull request #513
  • change: Merge pull request #514
  • change: Merge pull request #515
  • change: Merge pull request #516
  • change: Merge pull request #517
  • change: Merge pull request #518
  • change: Merge pull request #519
  • change: Merge pull request #520
  • …and 23 more

Architecture

  • Unchanged — 0 containers · 1 contexts · 0 edges

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

google-research/timesfm was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 15 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 918df02c9d366b2145bc5876135fa2131d3ed2a6 — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.73+preprod-52de90b74216.