In a recent development, Inception Labs, a startup founded by Stanford University professor Stefano Ermon, has unveiled its first product: Mercury Coder. This innovative tool is a large language model based on a diffusion mechanism (dLLM). Unlike conventional language models that generate text word by word, Mercury Coder leverages a diffusion-based approach to process entire text sequences simultaneously—a method akin to how AI generates images and videos.
Mercury Coder employs a coarse-to-fine generation strategy. It begins with a rough draft of the text and then refines it in parallel, similar to how tools like Midjourney and OpenAI’s Sora operate for image and video generation. This novel technique significantly boosts performance, reportedly achieving speeds up to 10 times faster than traditional models when running on NVIDIA H100 GPUs, producing over 1,000 words per second.
Performance benchmarks indicate that Mercury Coder matches or even surpasses leading models such as GPT-4o Mini and Claude 3.5 Haiku in certain scenarios. Additionally, its cost efficiency makes it an appealing choice for businesses aiming to optimize their AI infrastructure.
AI researcher Andrej Karpathy expressed interest in Mercury Coder's diffusion-based methodology, noting that this marks a departure from conventional practices in text generation. While diffusion methods have been widely used in image and video generation, they had not yet gained traction in text generation until now. This new model could uncover fresh advantages and limitations in AI-driven text creation.
Inception Labs currently offers both API access and local deployment options for Mercury Coder. The company has partnered with several Fortune 500 enterprises to help them reduce AI-related latency and costs. Moreover, hints suggest that more dLLM models optimized for conversational AI may be on the horizon.
Whether diffusion-based large language models can emerge as formidable competitors to traditional architectures remains to be seen. However, through Mercury Coder, Inception Labs has demonstrated that AI text generation isn’t confined to the sequential frameworks dominating the current landscape.
