mys/timesfm-3.0-GGUF

TimesFM 3.0 GGUF (ggmlc)

Google TimesFM 3.0 foundation time-series forecaster compiled from google/timesfm-3.0-pytorch (20 mixing layers, 1280 hidden, patch 32→64).

These files are not llama.cpp / llama-cli GGUFs. They are produced by ggmlc, a neural network compiler that lowers PyTorch / JAX / Flax / Keras models to high-performance GGML execution. Loading them in llama.cpp will fail.

Get running

1. Download a GGUF from this repo (see table below).
2. Download the pre-built timesfm binary from ggmlc GitHub Releases (or all releases).
3. CLI / serve / Web Studio details — see the example README: examples/timesfm (raw README).

RevIN, detrending, patching, quantile bands, and the Web Studio live in that standalone C++ binary — not in llama.cpp.

Files

File	Quant	Size (approx)	Notes
timesfm3_f16.gguf	F16	~632 MB	Reference / highest quality
timesfm3_q8_0.gguf	Q8_0	~336 MB	Strong doctor parity vs F16
timesfm3_ud_q4_k_m.gguf	UD_Q4_K_M	~304 MB	Unsloth-dynamic Q4_K_M; preferred 4-bit
timesfm3_q4_0.gguf	Q4_0	~179 MB	Smallest; larger forecast deviation vs F16 is expected for this forecaster (not an LLM logits-over-vocab setup). doctor may fail the weekly-retail sMAPE gate by a thin margin while F16/Q8/UD pass

huggingface-cli download mys/timesfm-3.0-GGUF timesfm3_f16.gguf --local-dir .

Quick start

Put the release timesfm binary on your PATH (or run it by path), then:

.\timesfm.exe info timesfm3_f16.gguf
.\timesfm.exe doctor timesfm3_ud_q4_k_m.gguf --device cuda
.\timesfm.exe forecast timesfm3_f16.gguf --preset weekly_retail --horizon 28 --non-negative --device cuda
.\timesfm.exe serve timesfm3_f16.gguf --port 8080

./timesfm info timesfm3_f16.gguf
./timesfm doctor timesfm3_ud_q4_k_m.gguf --device cuda
./timesfm forecast timesfm3_f16.gguf --preset weekly_retail --horizon 28 --non-negative --device cuda
./timesfm serve timesfm3_f16.gguf --port 8080

Notes

• Prefer F16 or Q8_0 when you need closest match to the PyTorch reference.
• UD_Q4_K_M is the recommended compact quant for everyday local use.
• Q4_0 is shipped for size-constrained deployments; expect larger point-forecast / sMAPE drift vs F16 on some presets (e.g. weekly retail). That is a property of aggressive 4-bit block quant on a continuous forecaster, not a ggmlc runtime bug.

License

Apache 2.0, same as google/timesfm-3.0-pytorch. Compiler: ggmlc (MIT).