Stability AI Launches 'Smol' Model: Introducing StableLM Zephyr 3B

Stability AI has gained recognition primarily for its innovative suite of text-to-image generative AI models, but the company is now expanding its offerings.

Today, Stability AI unveiled its latest advancement: StableLM Zephyr 3B. This model is a 3 billion parameter large language model (LLM) designed for various chat applications, including text generation, summarization, and content personalization. StableLM Zephyr 3B is an optimized, smaller version of the initial StableLM text generation model introduced earlier this year.

One of the key advantages of StableLM Zephyr 3B is its size. Being smaller than the 7 billion parameter versions of StableLM allows for deployment on a broader range of hardware with a reduced resource footprint, all while ensuring rapid responses. The model has been specifically optimized for question-answering and instructional tasks.

“StableLM was trained longer on higher quality data than previous models, utilizing twice the number of tokens compared to LLaMA v2 7b, yet it matches that model's base performance at just 40% of the size,” stated Emad Mostaque, CEO of Stability AI.

Introducing StableLM Zephyr 3B

StableLM Zephyr 3B is not an entirely new model but rather an extension of the existing StableLM 3B-4e1t model. Its design is informed by the Zephyr 7B model from HuggingFace, which operates under an open-source MIT license and is intended for use as an assistant. Zephyr employs a training method called Direct Preference Optimization (DPO), which is also utilized in StableLM.

Mostaque elaborated that DPO serves as an alternative to the reinforcement learning techniques used in earlier models, refining them to better align with human preferences. While DPO has generally been applied to larger models, StableLM Zephyr marks one of the first instances of this approach being used effectively in a smaller 3 billion parameter model.

Stability AI leveraged DPO alongside the UltraFeedback dataset from the OpenBMB research group, which comprises more than 64,000 prompts and 256,000 responses. This combination of DPO, the model's size, and the optimized training dataset results in impressive performance metrics. For example, in the MT Bench evaluation, StableLM Zephyr 3B outperformed larger models, including Meta's Llama-2-70b-chat and Anthropic's Claude-V1.

An Expanding Portfolio of Models

StableLM Zephyr 3B contributes to a growing array of models being released by Stability AI as the generative AI startup aims to enhance its capabilities continuously. In August, the company launched StableCode for application code development, followed by Stable Audio in September for text-to-audio generation. November saw the preview of Stable Video Diffusion, marking Stability AI's entry into video generation.

Despite this expansion, Stability AI remains committed to its roots in text-to-image generation. Recently, the company introduced SDXL Turbo, a faster version of its flagship SDXL text-to-image stable diffusion model.

Mostaque emphasizes that there is still much innovation to come from Stability AI. “We believe that small, open, performant models tailored to users' specific data will outperform larger general models,” he explained. “With the upcoming full release of our new StableLM models, we look forward to further democratizing generative language models.”