budibase Releases

This release introduces a new Data Synthesis Module (`oumi synth`) and Hyperparameter Tuning Module (`oumi tune`), alongside expanded support for AWS Bedrock and NERSC Perlmutter HPC clusters.

Nov 18, 2025

This release introduces support for the Qwen3-VL model and the NERSC Perlmutter HPC cluster. It includes various bug fixes related to GPU testing and Sky Pilot compatibility, alongside performance optimizations and code quality improvements.

Oct 20, 2025

This release enhances the Ultravox speech-language model by introducing new variants and improving audio robustness. It specifically adds support for Gemma3 and Qwen3 models while boosting Hindi language understanding. Additionally, the release integrates a new VoiceBench evaluation suite and improves noise handling capabilities.

Aug 18, 2025

Adds a configuration parameter to manage callback thread execution for better performance.

May 3, 2025

v0.3.103 introduces robust inter-process communication, asynchronous event handling, and audio normalization to improve the stability and metadata precision of the RealtimeSTT service.

Apr 19, 2025

v0.5 significantly boosts performance with a 60% increase in transcription accuracy and expands language support to 42 languages, alongside improvements in translation and question-answering capabilities.

Feb 11, 2025

v0.4.1 enhances model quality by upgrading the Whisper encoder to Whisper-large-v3-turbo and adding support for six new languages, resulting in better translation metrics across various language pairs.

Nov 12, 2024

Minimal release notes: "This includes the extended context size to 8192 so should provide 5 mins conversations on moshi-mlx."

Sep 22, 2024

Sep 20, 2024

v0.4 upgrades the Whisper encoder from small to medium and expands training data to support nine languages, resulting in significantly improved speech translation performance across various pairs.

Aug 27, 2024

v0.3 introduces improved speech understanding through Knowledge Distillation and synthetic data augmentation, utilizing a frozen Llama 3.1 8B core trained on 2.5k hours of audio.

Aug 23, 2024