jaichang/timesfm-2p5-200m-litert

TimesFM 2.5 200M LiteRT (TFLite) Variants

LiteRT/TFLite exports of google/timesfm-2.5-200m-pytorch for on-device inference.

Included files

• timesfm-2p5-200m-litert_ctx512_h128_fp32.tflite
• timesfm-2p5-200m-litert_ctx512_h128_fp16_w16a32.tflite
• timesfm-2p5-200m-litert_ctx512_h128_drq_w8a32.tflite
• timesfm-2p5-200m-litert_ctx512_h128_int4_dq_w4a4.tflite

Tensor spec

• Input: series, float32, shape [1, 512]
• Output: point forecast, float32, shape [1, 128]

Variant summary

File	Quantization	Notes
*_fp32.tflite	none	Best fidelity, largest size
*_fp16_w16a32.tflite	fp16 weights	Good speed/size balance on many devices
*_drq_w8a32.tflite	dynamic range quant (int8 weights, fp32 activations)	Smaller model, usually small quality drop
*_int4_dq_w4a4.tflite	dynamic int4 weight quant	Smallest model, accuracy/speed trade-offs depend on device

Source and license

• Base model: google/timesfm-2.5-200m-pytorch
• Original project: google-research/timesfm
• This repo redistributes converted variants of the Apache-2.0 base model.

Minimal Python inference example

import numpy as np
import tensorflow as tf

interpreter = tf.lite.Interpreter(
    model_path="timesfm-2p5-200m-litert_ctx512_h128_fp32.tflite"
)
interpreter.allocate_tensors()

inp = interpreter.get_input_details()[0]
out = interpreter.get_output_details()[0]

x = np.zeros((1, 512), dtype=np.float32)
interpreter.set_tensor(inp["index"], x)
interpreter.invoke()
y = interpreter.get_tensor(out["index"])
print(y.shape)  # (1, 128)

Notes

• These exports are intended for inference only.
• Validate numerics against your own ONNX/PyTorch baseline before production use.