Chronos-2 (Amazon) — GGUF

GGUF conversion of Amazon's Chronos-2 — an encoder-only bidirectional 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: F16.

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 amazon/chronos-2.

Context must be at least 32 timesteps (Chronos-2's patch_size is 16; shorter contexts are rejected outright rather than padded — 32 gives headroom) — 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/chronos-rs-gguf", "chronos-f16.gguf")
config_path = hf_hub_download("amazon/chronos-2", "config.json")

model = zsfm.ChronosModel(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]
point = model.forecast(context, horizon=64)
# -> List[float], the median (q0.5) forecast

Chronos has the richest Python surface of the single-checkpoint models here: besides forecast() (median), it also exposes forecast_quantiles(context, horizon) -> List[List[float]] (the full quantile matrix) and quantiles() -> List[float] (the quantile levels the matrix rows correspond to).

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original weights and converts to GGUF locally
# (produces the same bytes as chronos-f16.gguf in this repo) — `convert` also caches config.json exactly where `infer --config` defaults to, so it's omitted below:
zsfm chronos convert --dtype f16 -o gguf/chronos-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 chronos infer --gguf gguf/chronos-f16.gguf

-m/--model takes the full HuggingFace repo id (default amazon/chronos-2) — there's only one published checkpoint for this architecture, so you normally don't need to change it. -o/--output defaults to gguf/chronos-f16.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 chronos convert --dtype f32 -o gguf/chronos-f32.gguf
zsfm chronos convert --dtype q8  -o gguf/chronos-q8.gguf

To skip conversion and run a file already published here:

huggingface-cli download amaye15/chronos-rs-gguf chronos-f16.gguf --local-dir .
huggingface-cli download amazon/chronos-2 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], "horizon": 64}' \
  | zsfm chronos infer --gguf chronos-f16.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": "chronos",
  "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

Chronos-2 is an encoder-only bidirectional model:

• Input: Time series values are instance-normalized, patched, and concatenated with time encodings and observation masks
• Encoder: Alternating TimeSelfAttention + GroupSelfAttention + FeedForward blocks, all with T5-style RMSNorm
• RoPE: Standard Llama rotate_half ([-x[half:], x[:half]]), unlike Toto which uses xPos
• Attention: Scale = 1.0 (no 1/√d scaling, per the original implementation)
• Output: Last n_output_patches hidden states → ResidualBlock → quantile predictions

For batch=1 (univariate inference), GroupSelfAttention reduces to a position-wise v → o projection.

License

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