Open weights · public benchmarks
We release open models with the evidence needed to test every claim: weights, benchmarks, limitations, and reproducible results. No black boxes.
New · VoiceOpen Nepali voice cloning from a 30-second voice note.
Speech recognitionOpen 119M-parameter Nepali ASR, small enough for a CPU.
Phone-call speechNepali ASR for call-centers, IVR and code-switched calls.
A compact base model for architecture research.
New · Voice
Give Swar 20 to 30 seconds of anyone's voice, a WhatsApp voice note is enough, and it speaks for them: Nepali, English, and Nepali mixed with English. It has the lowest pronunciation error on NepTTS-Bench and is first on the Nepali voice cloning benchmark.
Speech recognition
Kriti V1 is a 119M-parameter Nepali-only speech recognition model. It shares rank one of 19 evaluated systems on the Nepali development benchmark, records the lowest raw word error rate of all of them, and was trained on 630 hours of open Nepali audio.
Phone-call speech
Phone audio is 8 kHz, compressed and noisy, and models trained on studio speech fall apart on it. Kriti Telephony is trained for real calls and sets a new state of the art on NepTel, the public benchmark of Nepali call-center recordings.
Language model
BarunLM is a 35M-parameter decoder-only base model, built to study how architecture, data quality and evaluation interact under hard compute limits. Twelve layers repeat three local-attention blocks and one global one. It is a measuring instrument, not a chatbot.
