Mercury 2

Mercury 2 is a frontier-scale "Reasoning-First" foundation model developed by Inception Labs. Released as a successor to their initial breakthrough architecture, it is designed to prioritize "System 2" thinking—deliberate, logical, and verifiable reasoning—over the rapid, pattern-matching "System 1" responses typical of standard LLMs. It is widely recognized for its massive 2-million token context window and its ability to solve complex, multi-dimensional problems that require a deep understanding of physical laws and logical constraints.

What It Is

Mercury 2 is a high-intelligence reasoning model built on a proprietary Neural-Symbolic hybrid architecture. Unlike standard transformers that predict the next most likely word, Mercury 2 is trained to build internal logical models of a problem before generating an answer. It is specifically engineered to eliminate the "lazy" reasoning often seen in AI, ensuring it follows through on every step of a complex derivation without skipping logic or hallucinating intermediate facts.

What It Can Do

• Massive Context Processing: Handles up to 2 million tokens, allowing it to analyze dozens of thick textbooks, entire legal archives, or massive software repositories in a single pass.
• Native Chain-of-Thought: Features a built-in "Thinking" process that it uses to self-correct and verify its logic in real-time before providing a final output.
• Spatial and Physical Reasoning: Demonstrates a "world model" capability, allowing it to predict physical outcomes, understand 3D spatial relationships, and solve complex engineering puzzles.
• Perfect Recall: Achieves "needle-in-a-haystack" retrieval accuracy across its entire 2M context window, making it nearly impossible for it to forget a specific detail buried in a massive dataset.
• High-Fidelity Tool Use: Can autonomously operate complex software interfaces, such as CAD tools, specialized scientific simulators, and terminal-based environments.

Examples of Its Capabilities

Mercury 2 is uniquely capable of handling "Long-Horizon Engineering," where it can ingest a 500-page technical manual for a complex piece of machinery and then autonomously design a compatible sub-system. For example, when tasked with designing a custom structural bracket, the model doesn't just sketch a shape; it performs the stress analysis calculations, selects materials based on thermal constraints, and generates the exact CAD files needed for production. Its 2-million token window allows it to reference every single constraint and tolerance mentioned in the source documentation, ensuring that the final output is not just a guess, but a mathematically verified solution.

In the legal and medical sectors, Mercury 2 acts as a "Super-Researcher" capable of synthesizing thousands of pages of case law or patient history into a single, cohesive strategy. It can track subtle contradictions across 20 different witness depositions or identify a rare drug interaction buried in years of unstructured medical notes. Because of its "Reasoning-First" nature, the model can explain exactly why it reached a specific conclusion, citing the precise page and paragraph from the source data, which significantly reduces the time required for human experts to verify its work.

How Does It Work?

Mercury 2 moves away from the pure "predict-the-next-token" paradigm by utilizing Dynamic Computation Allocation. This means the model can choose to "think" longer on a hard math problem than it does on a simple greeting, effectively spending more "brainpower" where it is needed most. It uses a Linear-Complexity Attention mechanism (likely based on advanced State Space Models or SSMs) which prevents the massive slowdowns that usually occur when processing very long documents. Its training involves Verifiable Reinforcement Learning, where the model is rewarded not just for the correct final answer, but for the logical soundness of every individual step taken to get there.

Applications of Mercury 2

The applications of Mercury 2 are primarily found in high-stakes environments such as autonomous engineering and CAD design, where precise spatial reasoning and physical law adherence are non-negotiable. It is increasingly used in complex legal discovery and medical research to synthesize millions of words of documentation without losing track of fine-grained details. Furthermore, its ability to manage massive codebases makes it a premier choice for legacy system migration, where it can map out the logic of millions of lines of old code and rewrite it into modern languages while maintaining perfect functional parity.

Previous Models

• Mercury 1 (Late 2024): The proof-of-concept model that first introduced Inception Labs' reasoning-centric approach; it featured a smaller 128K context window and was primarily used for coding benchmarks.‍
• Mercury 1.5 Pro (Early 2025): An intermediate release that scaled the context window to 1 million tokens and introduced the first iteration of the "Dynamic Computation" engine.