加载中…
加载中…
A Russian text-continuation model in 0.36 B parameters, converted to GGUF for llama.cpp. It continues a
picked prefix instead of answering it, so it suits keyboard suggestions and autocomplete, not chat.
Trained by continued pretraining on 1.2 B tokens (0.9 B ru + 0.204 B en + 0.096 B zh) on top of
taobao-mnn/MNN-Pocket-0.3B-Base. Weights, training recipe
and measurements live in Rahspide/MNN-Pocket-0.3B-RU.
llama-cli -hf Rahspide/MNN-Pocket-0.3B-RU-GGUF:Q4_K_M -p "Сегодня утром я" -n 40
| File | Size | Note |
|---|---|---|
| F16 | 581 MB | reference, not for shipping |
| Q8_0 | 310 MB | near-lossless, desktop |
| Q4_K_M | 223 MB | the size to ship on a phone |
Measured on a 16-thread x86 CPU: Q4_K_M ~123–130 tok/s generation, Q8_0 ~99–103 tok/s. Phone numbers are not measured here.
The tokenizer is byte-level, so a token can end in the middle of a Cyrillic character. llama.cpp buffers the
partial UTF-8 sequence itself; a hand-rolled decoder has to do the same.