ApsuaT9-v1
ApsuaT9-v1 is a multilingual word completion model for Abkhaz, Russian, English and Turkish. It suggests words from the preceding text and the letters already typed. Inference runs locally on the CPU.
Model
| Property | Value |
|---|---|
| Architecture | 4-layer language model |
| Vocabulary | 16,384 tokens |
| Maximum context | 128 tokens |
| Model format | ONNX |
| Weight storage | INT8, groups of 32 |
| Model file | 15.25 MB |
| Computation and KV cache | FP32 |
| Default search | Beam 5, up to 3 suggestions |
The weights and tokenizer are identical to those used by Apsua Studio. INT8 describes weight storage; inference uses FP32 operations.
Usage
pip install "git+https://huggingface.co/apsua/ApsuaT9-v1"
from apsua_t9 import T9
model = T9.from_pretrained("apsua/ApsuaT9-v1")
print(model.suggest("Сегодня мы по", language="ru"))
Pass the model some text to get suggested continuations. It can complete a partially typed word or suggest the next word. For long texts, the model uses the most recent context. Language can be ab, ru, en, or tr; if omitted, inference runs without a language tag. An empty list is returned when there are no suitable suggestions. On the first run the model is downloaded from Hugging Face and cached locally.
License
Model weights and tokenizer: CC BY 4.0. Runtime code: Apache-2.0.