Instructions to use mlboydaisuke/Timestral-209M-LiteRT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LiteRT
How to use mlboydaisuke/Timestral-209M-LiteRT with LiteRT:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 1,632 Bytes
11f5cc3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | Timestral-209M — LiteRT
Model: Timestral-209M by Mistral AI (huggingface.co/mistralai/Timestral-209M, revision
9874d252ab0aae16f7f7c999f98f98d6ad7d8365). The weights are licensed under the Modified MIT License
(LICENSE; attribution notice: 2026 - Mistral AI). Its condition 2 withholds every right under that
license from a user whose company, or employer, had a global consolidated monthly revenue above $20
million (or its equivalent in another currency) in the preceding month, and it applies to the model
and to any derivatives, modifications or combined works based on it. The graph files
(timestral-209m-*.tflite) and the host tables (forecast_query.f32, rope_cos.f32, rope_sin.f32) of this
repository are converted from these weights and are covered by that license.
The code files host_forecast.py, timestral_litert.py, android/TimestralHost.kt and android/GraphSpec.kt
are licensed under the Apache License 2.0 (LICENSE-APACHE-2.0). They port parts of the official
inference code (github.com/mistralai/timestral, commit a534bdbd244d64f3d9d743f46c7bf6fc330ed4ff, Apache
License 2.0): the causal scaler, the asinh normalization, the denormalization with clamp_nonfinite, the
advancing refine frame of _refine_block, the quantile cap, the NaN check, symmetric averaging and
non-negative clipping of TimestralForecaster.forecast, and the attention mask and positions of the
forecast's two token streams. Changes include one decoder block of at most 32 patches (1,024 steps), a
context of at most 8,192 steps, fixed graph shapes with left padding, and the series and its negation
computed as the two rows of one graph call.
|