Inception Labs has officially launched Mercury 2, a new type of inference language model positioned as the fastest in the world. Its innovative diffusion technique allows for parallel token generation, completely changing traditional AI text generation methods and significantly reducing response times in production deployments.
Technical Innovation: Breakthrough in Diffusion Model Architecture
Traditional large language models rely on an auto-regressive mechanism, generating tokens one by one. Mercury 2, however, utilizes diffusion technology to generate multiple tokens in parallel, progressively refining the output through a finite number of steps. This non-auto-regressive method not only accelerates the inference process but also optimizes hardware resource utilization. According to official benchmarks on an NVIDIA Blackwell GPU, the model reaches an inference speed of 1009 tokens/second, a more than 5x improvement over speed-optimized models.
Detailed Specifications
Mercury 2 offers a long 128K context window, supports native tool use, and has the capability for structured output that strictly adheres to predefined schemas. These features enable it to handle complex tasks. The pricing strategy is attractive: $0.25 per million input tokens and $0.75 per million output tokens. The model’s quality is described as on par with current top-tier speed-optimized models, while demonstrating a significant speed advantage—at least twice as fast as GPT-5.2.
Diverse Application Scenarios
In practical applications, Mercury 2 is particularly well-suited for coding assistance in software development, including code auto-completion, next-edit suggestions, code refactoring, and interactive AI coding agents. It can also be used to build efficient agentic workflows, enabling chained inference calls. For real-time interactions, the model supports voice interfaces and AI-driven video avatar generation. Furthermore, in information retrieval and RAG (Retrieval-Augmented Generation) systems, it can efficiently perform multi-hop searches, result re-ranking, and content summarization tasks.
Immediately Available and Ecosystem Compatible
Mercury 2 is now open to developers and can be accessed via the Inception Labs API platform. The platform provides an OpenAI-compatible interface for easy integration. Interested enterprise users can contact the company for custom evaluation support to validate the model’s suitability for specific workloads. The Inception Labs team, led by CEO Stefano Ermon, is working closely with leading infrastructure providers like NVIDIA, dedicated to providing high-performance solutions for all types of AI applications. Users can also experience Mercury Chat by visiting chat.inceptionlabs.ai.
