
Mercury 2 is the world's fastest reasoning language model, designed to make production AI feel instant. This large language model uses a revolutionary parallel refinement diffusion technology that generates multiple tokens simultaneously, achieving speeds exceeding 1,000 tokens per second. Developed for developers and enterprises who demand real-time AI performance, Mercury 2 redefines the speed-quality trade-off in language models. Its core value lies in eliminating the latency bottlenecks that plague traditional autoregressive models, enabling responsive AI applications that feel immediate to end users. By rethinking the decoding process, Mercury 2 brings diffusion-based reasoning to production workloads, setting a new standard for speed in AI inference. Built on a new foundation, it targets latency-sensitive use cases where every millisecond matters, from interactive coding assistants to real-time voice agents.
Current large language models rely on autoregressive sequential decoding, generating one token at a time from left to right. This creates a fundamental bottleneck in production AI systems, where multiple inference calls are chained together in loops. In agentic workflows, retrieval pipelines, and extraction jobs, latency compounds across every step, every user, and every retry, quickly making systems feel sluggish. Mercury 2 directly addresses this pain point with its diffusion-based approach. By generating responses through parallel refinement over a small number of steps, it achieves over five times faster generation than traditional models. This speed change is not just incremental; it fundamentally alters the feasibility of running reasoning-quality models in real-time applications, unlocking new possibilities for interactive AI without the usual latency trade-offs.
Mercury 2's standout feature is its unprecedented speed: 1,009 tokens per second on NVIDIA Blackwell GPUs. This is achieved through its parallel refinement diffusion technology, which produces multiple tokens simultaneously and converges to a high-quality response over just a few refinement steps. The benefit for users is immediate responsiveness in moments they experience—p95 latency under high concurrency remains low, throughput stays stable even when systems are busy. This speed not only enhances user experience but also allows developers to run more complex reasoning chains within the same latency budget. The model's speed curve is fundamentally different from autoregressive models, making production AI feel truly instant rather than a sequential wait. For real-time applications like code autocomplete or voice assistants, this means suggestions arrive fast enough to feel part of the user's own thinking, creating a seamless interactive experience.
admin
Despite its extreme speed, Mercury 2 does not compromise on reasoning quality. It offers tunable reasoning, allowing developers to adjust the depth of reasoning to suit their application's speed-quality requirements. The model supports a 128K context window, enabling it to process extensive documents and conversations in a single pass. Native tool use allows Mercury 2 to call external functions and APIs, extending its capabilities beyond text generation. Additionally, it can produce schema-aligned JSON output, ensuring structured data that integrates directly into downstream applications. These features collectively make Mercury 2 not just fast, but also intelligent and adaptable, suitable for complex production tasks that require both speed and precision. The combination of tunable reasoning and structured output empowers developers to build AI systems that are both responsive and reliable, reducing the need for post-processing and validation.
Mercury 2 is designed for seamless integration into existing stacks. It is fully OpenAI API compatible, meaning developers can drop it in without rewriting any code. This compatibility lowers the barrier to adoption and allows teams to upgrade their AI infrastructure with minimal friction. The pricing is highly competitive: $0.25 per million input tokens and $0.75 per million output tokens. When combined with its speed, this results in a very low cost per token under high throughput, making it economically viable for large-scale production deployments. For enterprise evaluations, Inception Labs partners with customers on workload fit, eval design, and performance validation under expected serving constraints, ensuring a successful deployment. The affordability and compatibility mean that even resource-constrained teams can harness world-class reasoning speed without breaking their budget or requiring architectural changes.
Mercury 2 is built on a diffusion-based language modeling architecture that fundamentally changes the generation process. Instead of the traditional typewriter approach of predicting one token at a time, it starts with a rough draft of the entire response and then refines it through multiple parallel diffusion steps. This process converges quickly to a coherent and accurate output, allowing the model to generate full responses in a fraction of the time. The diffusion method also alters the reasoning trade-off: higher intelligence typically requires more test-time compute, but diffusion enables reasoning-grade quality within real-time latency budgets. Mercury 2 optimizes for speed users actually feel, focusing on responsiveness in moments users experience, consistent turn-to-turn behavior, and stable throughput under load. This approach is particularly powerful for multi-step reasoning tasks where traditional models would take seconds; Mercury 2 can complete them in milliseconds, maintaining the illusion of instant intelligence.
In coding and editing, Mercury 2 powers autocompletions and next-edit suggestions that feel like part of the developer's own thinking, as noted by Zed co-founder Max Brunsfeld. For agentic loops, companies like Viant and Skyvern use Mercury 2 to optimize campaign execution and perform web tasks at scale, cutting latency per call to afford more reasoning steps. In real-time voice, Happyverse AI leverages Mercury 2 for lifelike video avatars that hold natural conversations, while OpenCall uses it for more responsive voice agents. SearchBlox integrates Mercury 2 into their search product to deliver sub-second intelligence across customer support, compliance, and e-commerce data. These use cases demonstrate that Mercury 2's speed translates directly into better user experiences and more efficient systems, enabling applications that were previously impractical.
Mercury 2 is ideal for developers and organizations building latency-sensitive AI applications. Specific roles include CTOs, chief product officers, and engineering leaders evaluating AI infrastructure. The model runs on NVIDIA Blackwell GPUs and is available now via an OpenAI-compatible API, with pricing at $0.25 per million input tokens and $0.75 per million output tokens. It caters to use cases spanning coding tools, autonomous advertising agents, voice interfaces, and enterprise search. In addition, Inception Labs offers enterprise evaluation support, partnering on workload fit and performance validation. With its combination of extreme speed, diffusion-based reasoning, and competitive pricing, Mercury 2 provides a transformative solution for production AI. It allows teams to deliver instant, intelligent responses without the latency pain of traditional LLMs, making real-time reasoning a practical reality for a wide range of industries. This model sets a new standard for what's possible in AI inference.
Mercury 2 is built for developers, CTOs, and product leaders at technology companies deploying AI-powered applications that require real-time responsiveness. Specific roles include engineering teams building coding assistants, autonomous agents, voice interfaces, and search engines. It is also ideal for enterprises evaluating AI models for production workloads where latency directly impacts user experience and operational efficiency. Industries such as software development, advertising technology, customer support, and human-computer interaction can benefit from Mercury 2's combination of speed and reasoning quality. The model runs on NVIDIA Blackwell GPUs and is accessible via an OpenAI-compatible API, making it suitable for teams already integrated into the OpenAI ecosystem.
Updated 2026-02-28