amaye15/timesfm-gguf

TimesFM 2.5 (Google) — GGUF

GGUF conversion of Google's TimesFM 2.5 — a 200M parameter decoder-only patch transformer. 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.

F32	F16	Q8_0
timesfm-f32.gguf	timesfm-f16.gguf	timesfm-q8.gguf

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

Context must be at least 32 timesteps (TimesFM patches the context in blocks of 32) — a shorter context fails with context too short. The examples below use a 32-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/timesfm-gguf", "timesfm-f16.gguf")

model = zsfm.TimesFmModel(gguf_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]
point = model.forecast(context, horizon=64)
# -> List[float], the point forecast (TimesFM's own dedicated point-forecast output, not a quantile)

# full 9-quantile distribution (added in zsfm 0.2.7)
quantile_matrix = model.forecast_quantiles(context, horizon=64)  # 9 rows, one per level below
levels = model.quantiles()  # [0.1, 0.2, ..., 0.9] — quantile_matrix[i] is the level[i] forecast
# note: quantile_matrix[4] (q0.5) is close to but NOT identical to `point` above —
# TimesFM decodes them as two separate output channels, not one derived from the other.

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original weights and converts to GGUF locally
# (produces the same bytes as timesfm-f16.gguf in this repo):
zsfm timesfm convert --dtype f16 -o gguf/timesfm-f16.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], "horizon": 64}' \
  | zsfm timesfm infer --gguf gguf/timesfm-f16.gguf

-m/--model takes the full HuggingFace repo id (default google/timesfm-2.5-200m-pytorch) — there's only one published checkpoint for this architecture, so you normally don't need to change it. -o/--output defaults to gguf/timesfm.gguf (no dtype suffix — unlike every other model here) 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 timesfm convert --dtype f32 -o gguf/timesfm-f32.gguf
zsfm timesfm convert --dtype q8  -o gguf/timesfm-q8.gguf

To skip conversion and run a file already published here:

huggingface-cli download amaye15/timesfm-gguf timesfm-f16.gguf --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], "horizon": 64}' \
  | zsfm timesfm infer --gguf timesfm-f16.gguf

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": "timesfm",
  "choices": [{
    "index": 0,
    "forecast": {
      "point": [2.1, 2.3, 2.5],
      "quantiles": {
        "0.10": [1.8, 2.0, 2.2],
        "0.50": [2.1, 2.3, 2.5],
        "0.90": [2.4, 2.6, 2.8]
      }
    },
    "finish_reason": "stop"
  }],
  "usage": {"context_length": 32, "forecast_length": 64}
}

point is the median (q0.5); all 9 quantile levels (q0.10–q0.90) are included.

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

Architecture

TimesFM 2.5 is a decoder-only patch transformer:

• Input: Context is split into 32-value patches; each patch is instance-normalized (RevIN with cumulative running statistics) and concatenated with an observation mask, producing 64-dim tokenizer inputs
• Backbone: 20-layer causal transformer, d_model=1280, 16 heads, head_dim=80, d_ff=1280; 200M parameters
• Attention: Sandwich norms (pre-norm → attention → post-norm + residual); fused QKV projection; per-dimension query scale 1.442695/√80 × softplus(param); QK RMSNorm after RoPE
• AR decoding: KV cache with 4-patch decode stride — O(n) instead of O(n²) re-forward per step
• Output: Each output patch decoded through a ResidualBlock to [n_patches, 128, 10]; index 0 is point forecast, indices 1–9 are quantiles q0.1–q0.9

License

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