Mercury 2

Product Introduction

1. Definition: Mercury 2 is a diffusion-based large language model (LLM) engineered for real-time reasoning. It replaces traditional autoregressive decoding with parallel refinement technology, enabling simultaneous token generation.
2. Core Value Proposition: It eliminates latency bottlenecks in production AI loops (e.g., agentic workflows, RAG pipelines) by delivering reasoning-grade quality at speeds exceeding 1,000 tokens/sec, making real-time AI interactions economically viable.

Main Features

1. Parallel Refinement:
   • How it works: Instead of sequential left-to-right decoding, Mercury 2 generates multiple tokens concurrently through iterative refinement. It converges responses over 10–20 diffusion steps (vs. 100+ steps in autoregressive models).
   • Technology: Leverages diffusion probabilistic models adapted for language tasks, optimized via NVIDIA Blackwell GPUs.
2. Ultra-Low Latency:
   • Achieves 1,009 tokens/sec throughput under high concurrency (p95 latency).
   • Maintains stable throughput during peak loads, critical for voice interfaces and interactive tools.
3. Tunable Reasoning & Tool Integration:
   • Supports dynamic reasoning adjustment (e.g., chain-of-thought complexity) via step control.
   • Features native tool use, 128K context windows, and schema-aligned JSON output for seamless API integration.

Problems Solved

1. Pain Point: Autoregressive decoding in conventional LLMs creates compounded latency in multi-step AI loops (e.g., agents, RAG), degrading user experience and limiting workflow complexity.
2. Target Audience:
   • Developer Tools Engineers (e.g., Zed’s code-suggestion workflows).
   • Voice AI Developers (e.g., Happyverse AI’s real-time avatars).
   • Enterprise Automation Teams (e.g., Skyvern’s document processing agents).
   • RAG Pipeline Architects (e.g., SearchBlox’s retrieval systems).
3. Use Cases:
   • Agentic Workflows: 20+ inference calls/task without latency penalties.
   • Real-Time Voice Interfaces: Sub-second responses matching natural speech cadence.
   • Interactive Coding: In-flow code edits/refactors (e.g., Zed).
   • Multi-Hop RAG: Reasoning-enhanced search within sub-second budgets.

Unique Advantages

1. Differentiation:
   • 5x faster than sequential models while matching quality benchmarks of speed-optimized LLMs (e.g., GPT-5.2).
   • $0.75/1M output tokens pricing undercuts competitors requiring equivalent compute for similar tasks.
2. Key Innovation:
   • First reasoning diffusion LLM: Applies image-diffusion principles to language, enabling parallel token generation. This shifts the quality-latency tradeoff curve, allowing complex reasoning in previously infeasible real-time scenarios.

Frequently Asked Questions (FAQ)

1. How does Mercury 2 achieve 1,000+ tokens/sec?
   By replacing sequential decoding with parallel refinement diffusion, generating tokens simultaneously across NVIDIA Blackwell GPUs.
2. Is Mercury 2 compatible with existing AI stacks?
   Yes, it’s OpenAI API-compatible, requiring no code rewrites for integration into current pipelines.
3. What applications benefit most from Mercury 2?
   Latency-sensitive use cases like voice AI, agentic loops, and real-time coding tools where >200ms delays break user immersion.
4. How does diffusion improve reasoning quality?
   Parallel refinement allows more compute per token within fixed latency budgets, enabling complex chain-of-thought reasoning traditionally limited by sequential bottlenecks.
5. What hardware optimizes Mercury 2 performance?
   NVIDIA Blackwell GPUs deliver peak throughput, though the model runs efficiently on standard cloud infrastructure.