language, license, datasets, library_name, pipeline_tag, base_model, tags
| language | license | datasets | library_name | pipeline_tag | base_model | tags | |||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
cc-by-nc-4.0 |
|
gguf | text-generation | inspirebek/qwen3-4b-uzbek-v2 |
|
qwen3-4b-uzbek-v2-gguf
gguf suite for inspirebek/qwen3-4b-uzbek-v2. cpu / apple silicon / vulkan / rocm via llama.cpp, ollama, lm studio, etc.
files
| quant | size | notes |
|---|---|---|
f16 |
8.8 gb | reference fp16 |
Q8_0 |
4.7 gb | near-lossless |
Q6_K |
3.6 gb | recommended for quality |
Q5_K_M |
3.2 gb | balanced |
Q5_K_S |
3.1 gb | slightly lighter |
Q4_K_M |
2.7 gb | recommended for most users |
Q4_K_S |
2.6 gb | smaller, slight quality loss |
Q3_K_M |
2.2 gb | aggressive |
Q2_K |
1.8 gb | edge / low-ram only |
usage
llama.cpp:
llama-cli -m qwen3-4b-uzbek-v2-q4_k_m.gguf -p "Salom! Qalaysan?" -cnv
ollama:
ollama run hf.co/inspirebek/qwen3-4b-uzbek-v2-GGUF:Q4_K_M
quantization
converted from the bf16 merged model via llama.cpp's convert_hf_to_gguf.py → llama-quantize. no calibration data (k-quants are statistics-only).
datasets
stage a — fluency (continued pretraining):
yakhyo/uz-wiki· MITtahrirchi/uz-books-v2· MITtahrirchi/uz-crawl· Apache-2.0
stage b — instruct (sft):
saillab/alpaca_uzbek_taco· CC-BY-NC-4.0behbudiy/alpaca-cleaned-uz· CC-BY-4.0UAzimov/uzbek-instruct-llm· Apache-2.0CohereLabs/aya_collection_language_split· Apache-2.0med-alex/qa_mt_ru_to_uzn· unspecifiedmed-alex/qa_mt_tr_to_uzn· unspecified
⚠️ licensing note:
saillab/alpaca_uzbek_tacois cc-by-nc-4.0, which restricts commercial use of derivative models. downstream users who need a fully permissive license should retrain without that subset.