FlashLabs open-sources Chroma 1.0 for real-time voice chat
FlashLabs just open-sourced Chroma 1.0, a 4B end-to-end spoken dialogue model that directly understands audio inputs and replies with natural speech in real time. Built on Qwen2.5-Omni and Llama 3 under an Apache-2.0 license, it skips traditional TTS pipelines and supports instant voice cloning with a short audio prompt. It currently supports English and is hosted on Hugging Face.
FlashLabs Chroma 1.0 is an open-source, 4B-parameter multimodal model built for real-time, end-to-end spoken dialogue. Instead of stitching together separate speech-to-text, LLM, and text-to-speech tools, it handles voice conversations directly from audio in to audio out.
What it can do
Chroma 1.0 operates as a causal language model that natively processes auditory inputs. When a user speaks, it understands the audio stream directly and generates coherent text and synthesized speech at the same time.
It also comes with personalized voice cloning built in. By passing a reference audio clip alongside prompt text during inference, you can instruct the model to mirror that voice's timbre and tone in its vocal responses.
FlashLabs released the model weights and inference code under a permissive Apache-2.0 license, making it fully available for commercial experimentation and research.
Under the hood and getting started
The architecture combines several proven open components. Its reasoning engine is based on Qwen2.5-Omni-3B, connected to a 16-layer Llama 3 backbone and a 4-layer Llama 3 decoder. For audio compression and reconstruction, it relies on Kyutai's Mimi codec running at a 24kHz sampling rate.
Running Chroma 1.0 locally is straightforward. The weights are hosted on Hugging Face under FlashLabs/Chroma-4B, alongside ready-to-run Google Colab and Kaggle notebooks.
Developers only need recent versions of PyTorch and Transformers. The model loads through AutoModelForCausalLM and AutoProcessor, outputting raw tokens that the integrated codec decodes directly into 24kHz audio values.
Why it matters
Cascaded voice pipelines with separate ASR and TTS stages are quickly becoming obsolete for conversational agents because they introduce latency and strip out emotional nuance. Chroma 1.0 brings a complete, lightweight 4B end-to-end voice engine to the open-source community.
Keep in mind that it currently only supports English, so developers targeting multilingual setups will have to wait for fine-tunes. But for anyone building real-time voice agents or exploring personalized virtual companions right now, it offers a remarkably accessible starting point.
Sources (1)
- huggingface.co(opens in a new tab)Originalhuggingface.co