BLOOMChat is SambaNova’s 176-billion-parameter multilingual chat model, and this repository contains the complete code behind it: data preparation, tokenization, training and inference. The model is instruction-tuned from BLOOM-176B on assistant-style conversation data, bringing chat behavior to one of the largest openly released language models. The repository has about 580 stars on GitHub and was updated in August 2026.
BLOOMChat supports conversation, question answering and generative answers across multiple languages, reflecting BLOOM’s multilingual training. Its training data originated from the OIG collection of instruction-style datasets, and the fine-tuning was performed on SambaNova DataScale systems using the company’s Reconfigurable Dataflow Architecture, an alternative to GPU clusters for very large models. The trained weights, BLOOMChat-176B-v1, are published on Hugging Face under the SambaNova Systems organization. For researchers, the value of this repository is the end-to-end recipe: how a base model of this scale was prepared, tokenized, instruction-tuned and served, in code rather than only in a paper.
As one of the few openly documented instruction-tuning efforts at the 176-billion-parameter scale, the repository is frequently referenced by teams planning their own large fine-tunes. It also illustrates an alternative hardware path, since the run was executed on SambaNova’s dataflow systems rather than conventional GPU clusters, and the code shows how the data pipelines were organized for that environment.
The BLOOMChat code is publicly available on GitHub and the model weights are published on Hugging Face; check the repository and model card for the exact license terms before commercial use.
BLOOMChat is for researchers and engineers studying very large multilingual chat models, BLOOM fine-tuning, or alternatives to GPU-based training infrastructure.