TTM-R2 (IBM Granite TinyTimeMixer) — GGUF

GGUF conversion of IBM Granite's TTM-R2 (TinyTimeMixer) — a ~1M parameter compact time-series foundation model. Converted and run with zsfm, a Rust workspace that ports zero-shot forecasting and tabular foundation models to GGUF + candle. No PyTorch, no Python runtime required to run inference.

F16 is generally the best size/accuracy trade-off; Q8_0 is smallest. This repo's default recommendation matches the upstream conversion default: F32.

There's no config.json in this repo — GGUF embeds its own architecture metadata for the CLI, but the Python bindings still need config.json from ibm-granite/granite-timeseries-ttm-r2.

Context must be at least 64 timesteps (TTM's patch_length is 64; shorter contexts are rejected outright rather than padded) — a shorter context fails with context too short. The examples below use a 64-value context.

Use it

Python (pip install zsfm)

pip install zsfm huggingface_hub

import zsfm
from huggingface_hub import hf_hub_download

gguf_path = hf_hub_download("amaye15/ttm-gguf", "ttm-f32.gguf")
config_path = hf_hub_download("ibm-granite/granite-timeseries-ttm-r2", "config.json")

model = zsfm.TtmModel(gguf_path, config_path)

context = [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93, 2.01, 2.09, 2.17, 1.9, 1.98, 2.06, 2.14, 2.22, 2.3, 2.38, 2.11, 2.19, 2.27, 2.35, 2.43, 2.51, 2.59, 2.32, 2.4, 2.48, 2.56, 2.64, 2.72, 2.8, 2.53, 2.61, 2.69, 2.77, 2.85, 2.93, 3.01, 2.74]
point = model.forecast(context, horizon=64)
# -> List[float], a plain point forecast (median only; TTM's decoder has no quantile head, so there's nothing else to return)

config has no working default — pass it explicitly (an omitted config falls back to a hardcoded path that almost never exists on your machine, raising FileNotFoundError).

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original weights and converts to GGUF locally
# (produces the same bytes as ttm-f32.gguf in this repo) — `convert` also caches config.json exactly where `infer --config` defaults to, so it's omitted below:
zsfm ttm convert --dtype f32 -o gguf/ttm-f32.gguf
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93, 2.01, 2.09, 2.17, 1.9, 1.98, 2.06, 2.14, 2.22, 2.3, 2.38, 2.11, 2.19, 2.27, 2.35, 2.43, 2.51, 2.59, 2.32, 2.4, 2.48, 2.56, 2.64, 2.72, 2.8, 2.53, 2.61, 2.69, 2.77, 2.85, 2.93, 3.01, 2.74], "horizon": 64}' \
  | zsfm ttm infer --gguf gguf/ttm-f32.gguf

-m/--model takes the full HuggingFace repo id (default ibm-granite/granite-timeseries-ttm-r2) — there's only one published checkpoint for this architecture, so you normally don't need to change it. -o/--output defaults to gguf/ttm-f32.gguf regardless of --dtype, so always pass -o explicitly (as above) — otherwise repeated runs overwrite the same file under a name that may not even match the dtype you chose:

zsfm ttm convert --dtype f32 -o gguf/ttm-f32.gguf
zsfm ttm convert --dtype q8  -o gguf/ttm-q8.gguf

To skip conversion and run a file already published here:

huggingface-cli download amaye15/ttm-gguf ttm-f32.gguf --local-dir .
huggingface-cli download ibm-granite/granite-timeseries-ttm-r2 config.json --local-dir .
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93, 2.01, 2.09, 2.17, 1.9, 1.98, 2.06, 2.14, 2.22, 2.3, 2.38, 2.11, 2.19, 2.27, 2.35, 2.43, 2.51, 2.59, 2.32, 2.4, 2.48, 2.56, 2.64, 2.72, 2.8, 2.53, 2.61, 2.69, 2.77, 2.85, 2.93, 3.01, 2.74], "horizon": 64}' \
  | zsfm ttm infer --gguf ttm-f32.gguf --config config.json

Source, the other 9 time-series forecasters + 5 tabular models, and full docs: amaye15/zsfm-rs.

Response format

{
  "id": "forecast-000001932b7a1234",
  "object": "forecast",
  "created": 1736290000,
  "model": "ttm",
  "choices": [{
    "index": 0,
    "forecast": {
      "point": [2.1, 2.3, 2.5],
      "quantiles": {}
    },
    "finish_reason": "stop"
  }],
  "usage": {"context_length": 64, "forecast_length": 64}
}

This architecture has no quantile head, so quantiles is always empty and point is the only forecast.

Pass a batch of series ("context": [[...], [...]]) for one choice per series.

Architecture

TTM (Tiny Time-Mixer) is a compact encoder-decoder model:

• Encoder: Multi-layer mixer blocks operating across the patch dimension; adaptive patching levels allow different temporal resolutions simultaneously
• Decoder: Lightweight projection from encoder representations to the forecast horizon
• Scale: ~1M parameters — orders of magnitude smaller than transformer-based foundation models, competitive on short-horizon benchmarks
• Config: context_length, prediction_length, patch_length, patch_stride, d_model, num_layers, decoder_num_layers, and adaptive_patching_levels are loaded from config.json

License

Conversion code: MIT (amaye15/zsfm-rs). Weights: Apache-2.0, per IBM Granite's original release — unrestricted, including commercial use.