balbal-v2: CPU-friendly Turkish ASR models trained for under $250

On 2026-11-03, a developer posted to Reddit's LocalLLaMA community about balbal-v2, a Turkish-language ASR model representing the v2 release after a period of work on Turkish speech recognition. The release comprises two CTC + KenLM models built for CPU deployment, with each trained for under $250. The larger variant, balbal-v2-manas, has 313M parameters and handles both Turkish and English, achieving roughly 123x real time on 8 threads. The smaller balbal-v2 has 45M parameters and handles Turkish only, reaching approximately 236x real time. According to the post, even a single core on an old Celeron POS box running Windows 7 runs the model faster than real time. On Mu-Bench TR, a Turkish benchmark made up of 846 real customer-service calls, results from the official harness place Google Chirp 3 first by WER, with balbal-v2 second. No further developments appear in the available material, so the release stands at the above model configurations and benchmark figures.