mlx-timesfm is an independent, inference-only implementation of TimesFM 3.0 for Apple MLX. It loads the original 330.7M-parameter FP32 checkpoint directly, without converted weights or a PyTorch runtime, and runs forecasting on Apple silicon through MLX and Metal.

Features

• Native MLX model, transformer, normalization, masking, and forecast decode.
• Direct mx.load() of the original model.safetensors checkpoint.
• Univariate and multivariate forecasts with nine quantiles.
• Past-only and past/future dynamic covariates.
• Linear detrending, stitching, contiguous patch masking, and iterative RevIN.
• Strict validation of all 445 checkpoint tensors and their shapes.
• MLX-only CPU/GPU golden regression against a versioned real-data oracle.
• Full-FP32 matrix kernels by default for numerically stable inference.

Status

This is an initial alpha release. The reference inference path is implemented and parity-tested, but broader dataset validation and performance tuning remain ongoing.

Requirements

• Apple silicon Mac with Metal support
• Python 3.13
• MLX 0.32.x
• A separately downloaded TimesFM 3.0 checkpoint

Installation

Install from PyPI with uv:

uv add mlx-timesfm

Or install the current checkout for development:

git clone https://github.com/appautomaton/mlx-timesfm.git
cd mlx-timesfm
uv sync --locked

Model weights

Review and accept Google's checkpoint terms, then download the original files. One option is the Hugging Face CLI in an isolated uv tool environment:

uvx --from huggingface-hub hf download google/timesfm-3.0-pytorch \
  --local-dir models/timesfm_3_0/original

The checkpoint directory must contain:

models/timesfm_3_0/original/
├── config.json
└── model.safetensors

Weights remain local and are excluded from Git and all package distributions.

Quick start

import numpy as np

import mlx_timesfm

model = mlx_timesfm.load("models/timesfm_3_0/original")

# A batch of two univariate series, each with 512 context points.
target = np.random.default_rng(0).normal(size=(2, 512)).astype(np.float32)
quantiles = model.forecast(target, horizon=128)

print(quantiles.shape)  # (2, 1, 128, 9)
median = quantiles[..., 4]

Inputs may be one-dimensional (time), two-dimensional (batch, time), or three-dimensional (batch, variate, time). The output is always (batch, variate, horizon, 9), ordered from the 0.1 through 0.9 quantiles.

Covariates

target = np.zeros((1, 3, 512), dtype=np.float32)
past_only = np.zeros((1, 2, 512), dtype=np.float32)
past_future = np.zeros((1, 2, 512 + 128), dtype=np.float32)

quantiles = model.forecast(
    target,
    horizon=128,
    past_only_covariates=past_only,
    past_future_covariates=past_future,
)

Boolean masks use True for missing or invalid values. forecast() accepts a global context mask; decode() exposes separate masks for targets and both covariate groups.

Numerical precision

FP32 is the supported correctness baseline. MLX can otherwise choose reduced-precision matrix kernels while keeping FP32 array dtypes, so importing mlx_timesfm defaults MLX_ENABLE_TF32 to 0 before MLX computation. An explicit environment setting is preserved:

# Optional faster/reduced-precision experiment; not the parity baseline.
MLX_ENABLE_TF32=1 uv run python your_forecast.py

The model also pins the effective FP32 RMSNorm epsilon used to establish the golden oracle. FP16 and BF16 inference are not currently supported profiles.

Parity

The permanent regression uses 512 real context observations and a 128-step forecast from the UCI Appliances Energy Prediction dataset. It exercises three targets, two past-only covariates, two past/future covariates, every forecast quantile, and quantile ordering.

The golden manifest pins the dataset, source revision, checkpoint/config hashes, execution profile, output shape, and tolerance. These numbers establish implementation parity; they are not claims about forecast accuracy on every downstream dataset.

Development

uv sync --locked
uv run ruff check .
uv run pytest
uv build --no-sources

The normal test suite is MLX-only. With the original checkpoint present it runs the golden on MLX CPU and GPU; without the checkpoint, only the seven weight-dependent checks skip. A permanent hygiene test rejects retired framework imports and dependencies.

Release process

GitHub releases with tags matching v<version> trigger .github/workflows/workflow.yml. The workflow validates tag/version agreement, tests the project, builds and inspects both distributions, smoke-tests installed artifacts, and publishes through PyPI Trusted Publishing. Releases are not published from developer-machine credentials.

The pending PyPI publisher must match:

• Owner: appautomaton
• Repository: mlx-timesfm
• Workflow: workflow.yml
• Environment: pypi

Licensing

Project source code is Apache-2.0 licensed; see LICENSE. Third-party attributions are recorded in THIRD_PARTY_NOTICES.md. The separately distributed TimesFM 3.0 checkpoint has different, restrictive terms described above.

TimesFM is a Google trademark. This independent project is not affiliated with or endorsed by Google.

Built and maintained by AppAutomaton. Explore more MLX-native projects on GitHub and Hugging Face.