The ufakzeka family
Releases
Every model and benchmark of the family is listed here, newest first.
ufakzeka-kulak
It is a Turkish speech recognition model of 31,564,935 parameters, trained from scratch. It writes Turkish speech out as text with capitals and punctuation, writes numbers as words and runs offline on a CPU. It was trained on 4,977 hours of speech whose licences allow commercial use, and the same weights are released for NeMo, transformers, sherpa-onnx, parakeet.cpp (GGUF) and parakeet-mlx (MLX). In its release gate the word error rate is 0.0921 on FLEURS Turkish test (95 percent confidence interval 0.0792 to 0.1055) and 0.0916 on Common Voice 27.0 Turkish test (0.0873 to 0.0958), and it beat Whisper base on both; Whisper large-v3 scores lower, 0.0604 and 0.0767. It is weak on conversational speech and is not a streaming model. Measured limits are in the model card and on the model page.
ufakzeka-karar
It is a Turkish decision model of 182,494,466 parameters built on ufakzeka-1-base; the release is the last of three runs scored on HakemBench, and the research post tells how it got there. It reads a text and answers choice, score and yes/no questions in one forward pass per question, with a probability for every option and one not-sure signal per question. It ranks 7th of 16 rows on the HakemBench board (composite 0.660). Its numbers are not blind. Earlier runs’ test results shaped its training data, so its guardrail, moderation and support numbers are flagged; with every model scored on the other four tracks only, its composite is 0.678, 6th of 16. It is weak on support questions. The model is released in fp32, because the int8 files built from an earlier checkpoint did not pass the release gate. Measured limits are in the model card and on the model page.
HakemBench 1.0
This is the first version of the Turkish typed-decision benchmark. It was frozen on 26 September 2026 and is released fully open, with 2,346 items in seven tracks, 4,275 questions and four robustness probes. How the gold was made is on the benchmark page.
ufakzeka-1
151M parameters, Turkish from scratch. Base and chat models, f16 and q8_0 GGUF, nine release gates and hand-test counts in the model card. Post-training fixed safety phrasings, borrows through zero and percentages. Identity drift, multi-turn arithmetic and holding a refusal under pressure did not move, and all three are listed in the card with their numbers.