Stability AI has built a reputation for pushing the boundaries of artificial intelligence, and their latest release, Stable LM 2 1.6B, is no exception. This model is the latest addition to their stable of language models, which previously included 3 billion and 7 billion parameter models. The introduction of this more compact version shows Stability AI’s commitment to making AI more accessible and efficient.

Multilingual and efficient

Stable LM 2 1.6B model supports seven languages, namely English, Spanish, German, Italian, French, Portuguese, and Dutch. This expansion lets the model fit a wider audience and opens up new possibilities for cross-cultural communication and understanding.

Stable LM 2 1.6B finds the right balance between speed and performance, thanks to recent algorithmic advancements in language modeling. This balance means the model can handle a wide range of tasks without compromising on its ability to provide accurate and relevant responses in a timely manner.

Performance superiority

Despite its smaller size, when compared to the previous models, Stable LM 2 1.6B performs outstandingly. The new model reportedly outperforms other small language models with fewer than 2 billion parameters on most benchmarks. What’s even more remarkable is that it surpasses some larger models, including Stability AI’s own earlier Stable LM 3B model.

Drawbacks of smaller size

Stable LM 2 1.6B’s smaller size can bring a higher rate of hallucination. Hallucination in AI refers to the generation of content that is not factually accurate or contextually relevant. This can sometimes lead to responses that contain misleading or erroneous information.

There is also the potential for the model to generate toxic language that comes from smaller language models. Toxic language is language that is offensive, harmful, or inappropriate. Smaller, low-capacity language models may struggle to filter out such content, posing a challenge for developers and users.

Transparency and data utilization

Stability AI place a strong emphasis on the use of more diverse and extensive data in training Stable LM 2 1.6B. This includes documents in six languages besides English, making sure that the model is exposed to a wide range of linguistic patterns and cultural nuances.

This training process takes into account the order in which data is presented to the model, which helps the model better understand the contextual relationships between words and phrases.

Innovative training approach

One of the most exciting aspects of Stable LM 2 1.6B is the approach to training used by Stability AI. The company offers the new model in multiple formats, including pre-trained and fine-tuned versions, or the unique “last model checkpoint before the pre-training cooldown.”

This unique format lets developers take the model and further specialize it for specific tasks or datasets. This approach gives developers the chance to use the full potential of Stable LM 2 1.6B and tailor it to their specific needs.

Goal of the new model

So, what is the ultimate goal of Stability AI with the release of Stable LM 2 1.6B? The company’s vision is to provide developers with more tools and artifacts that they can use to innovate and build upon the current model.

By doing so, Stability AI aims to use the model’s capabilities in new and surprising ways. They want to catalyze a wave of creativity and problem-solving in the AI community, inspiring developers to explore uncharted territory and push the boundaries of what is possible with artificial intelligence.

Alexander Procter

February 5, 2024

3 Min