Built and shared.
For everyone.

The original LumiChats base model — a 3-billion parameter text generation model built for conversational AI. Designed for general-purpose chat with a focus on helpful, accurate, and context-aware responses.

A code-focused fine-tune of the LumiChats base, optimised for code generation, debugging, and explanation across popular languages including Python, JavaScript, and TypeScript.

A quantised 4-bit version of LumiChats v1.2 — an 8B model compressed for efficient inference on consumer hardware without significant quality loss. Ideal for running locally with limited VRAM.

LumiChats' first multimodal model — an 11-billion parameter vision model capable of understanding and describing images. Supports visual question answering, image captioning, and document understanding from screenshots.

An optimised 4-billion parameter model variant focusing on improved instruction following and task completion. Fine-tuned with a curated dataset to improve helpfulness and reduce hallucination compared to v1.1.

LumiChats' base text-to-speech model — converts written text to natural-sounding audio. Designed for Indian-accented English output and built as the foundation for the TTS fine-tune variants.

A fine-tuned variant of the TTS 1B base model, trained on a 16K-sample dataset for improved prosody, naturalness, and pronunciation accuracy — particularly for technical and academic vocabulary.

A 4-bit quantised fine-tune of the TTS base model, optimised for efficient inference on consumer hardware. Retains most of the speech quality of the full model while reducing memory requirements significantly.