TimesFM 3.0 merges native MLX backend: 641 series/sec on M4 Max, no PyTorch needed

rachittshah · x · 2026-09-08

A community PR adding an MLX backend to Google's official timesfm repo has been merged, with full feature parity with the PyTorch implementation.

• TimesFM 3.0 is Google's 330M-parameter foundation model for time-series forecasting, reading history in 32-point patches and writing the next 64 steps in a single forward pass (9 quantiles per step).
• The new backend lives in src/timesfm3/mlx with the same timesfm3forecaster interface; the torch code moved to src/timesfm3/torch with lazy re-exports, so import timesfm3 no longer pulls in PyTorch.
• M4 Max benchmarks (512-point context, 64-step horizon): 12.5 ms per series at batch 1; 641 series/sec at batch 32. Max deviation from the PyTorch reference is 2.1e-6 — float32 noise.
• The 1.9x speedup at batch 1 comes from mx.compile, not quantization: at 330M params the model is dispatch-bound, so bf16 and int8 yield little gain.

Original post →

More from Infra

• Arm's CSS for Mobile 2 Skips the NPU, Bets On-Device AI on CPU and GPU Engines — ryanshrout · 2026-09-08
• fal extends 75% off H3 Max endpoints to Sept 15, adds real-time 1080p video — isidentical · 2026-09-08
• CPU Shortage Reaches Software Teams as AI Chip Supply Constraints Spread Beyond AI — brada · 2026-09-08
• Big Tech scouts Argentina's Patagonia for new mega data centers — Polymarket · 2026-09-08
• Data center buildout is driving freight demand that the Cass index misses — kernelangus420 · 2026-09-08
• Strix Halo users ditch official llama.cpp: optimized forks hit ~60 t/s decode vs ~20 t/s — feelspeaceman · 2026-09-08

Infra channel →