Mercury 3

Product Introduction

1. Definition: Mercury 3 is a state-of-the-art, large-scale text diffusion language model (LLM) developed by Inception Labs. It represents a significant advancement in the diffusion-based AI model category, designed for high-performance, low-latency inference.
2. Core Value Proposition: Mercury 3 exists to prove that diffusion-based LLMs can scale to match or exceed the quality of leading autoregressive models (like those from OpenAI and Anthropic) while delivering dramatically higher inference speeds and lower operational costs. Its primary value is enabling real-time, latency-sensitive AI applications that were previously bottlenecked by model speed or cost.

Main Features

1. Diffusion at Scale Architecture: Mercury 3 utilizes a text diffusion architecture, a paradigm distinct from the standard autoregressive (next-token-prediction) approach used by most LLMs. This architecture allows for parallel token generation during inference, which is the core technical reason for its high speed. The model has been scaled to over 40 times the size of its predecessor, Mercury 1, demonstrating effective scaling laws for diffusion models.
2. High-Speed Inference Engine: The model is optimized for extreme throughput, achieving 1,068 tokens per second on standard NVIDIA GPUs. This is 4–8x faster than comparable frontier models like GPT-6 Luna and Claude Haiku 5.5. The technology focuses on minimizing Time-To-First-Answer-Token (TTFAT) and end-to-end latency, not just raw token generation.
3. Extended Context & Cost-Efficiency: Mercury 3 supports a 1-million-token context window, a substantial increase from the 260K in Mercury 2.5. It is offered at a competitive price of $0.20 per million input tokens and $0.70 per million output tokens, with launch discounts. The model is fully OpenAI API compatible, allowing for seamless integration into existing developer stacks and workflows.

Problems Solved

1. Pain Point: The high latency and cost of running powerful, frontier-level AI models in real-time, interactive applications. Slow response times break user immersion in applications like voice assistants, live tutoring, and autonomous agents.
2. Target Audience: AI product engineers and developers building latency-sensitive applications; enterprises in customer support, edtech, and search needing fast, accurate AI; researchers and companies evaluating diffusion model scalability.
3. Use Cases:
   • Autonomous Web Agents: For browser automation, research, and navigation where each agent step requires a model call, and delays compound.
   • Interactive Education & Tutoring: Powering real-time voice tutors and adaptive learning platforms that require instant feedback to maintain student engagement.
   • Real-Time Search and RAG: Processing dozens of documents in retrieval-augmented generation (RAG) pipelines fast enough to provide answers while the user is still engaged.
   • Live Customer Support: Driving chat and voice support agents that need to be both fast and highly accurate, especially in regulated industries.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros:
     • Unmatched Speed-to-Quality Ratio: Delivers frontier-model quality (benchmarked against GPT-6 Luna, Claude Haiku 5.5) at 4-8x the token generation speed.
     • Cost-Effective Inference: Lower per-token pricing combined with higher speed reduces total cost of ownership for high-volume applications.
     • Proven Scalability: Demonstrates that diffusion LLMs follow beneficial scaling laws, challenging the dominance of autoregressive architectures.
     • Developer-Friendly: OpenAI API compatibility ensures easy integration and migration.
   • Cons:
     • Architectural Novelty: The diffusion paradigm is less familiar to most developers than autoregressive models, potentially requiring a learning curve for fine-tuning and optimization.
     • Ecosystem Maturity: While growing, the tooling and community around diffusion LLMs are not as extensive as for autoregressive models like GPT.
     • Limited Public Model Details: As a proprietary commercial model, full architectural details, training data, and extensive benchmark results are not open-sourced for independent verification.

2. Key Alternatives & Differentiation:

   • OpenAI GPT-6 Luna / o1 Series: The benchmark for reasoning quality. Mercury 3 differentiates by being significantly faster and cheaper per token, making it more suitable for real-time, high-volume applications where ultra-fast reasoning is the priority over maximum depth of chain-of-thought.
   • Anthropic Claude Haiku 5.5: Known for its speed and strong performance. Mercury 3 claims to outperform it on quality benchmarks while matching or exceeding its speed, offering a direct alternative in the "fast, capable model" category.
   • Jev (Decision Models): Compared to specialized decision/classification models like Jev, Inception offers Mercury Decide, a tailored model within its portfolio. Mercury 3 itself is a general-purpose model, whereas Jev is optimized for high-volume classification tasks.

Frequently Asked Questions (FAQ)

1. What is a diffusion LLM and how is it different from ChatGPT? A diffusion LLM, like Mercury 3, generates text by iteratively refining random noise into coherent output, allowing parallel token prediction. ChatGPT uses an autoregressive method, predicting one token at a time sequentially. This fundamental difference is what enables diffusion models to achieve much higher inference speeds.
2. Is Mercury 3 really better than GPT-6? According to Inception Labs' benchmarks, Mercury 3 outperforms the speed-optimized variant GPT-6 Luna on a range of tasks and matches or exceeds its quality while being 4-8x faster. It is positioned as a superior choice for applications where latency is critical.
3. How much does the Mercury 3 API cost? The standard price is $0.20 per million input tokens and $0.70 per million output tokens. At launch, it is offered at a 50% discount: $0.10/M input and $0.35/M output tokens. Cached tokens receive a further 90% discount.
4. What are the main use cases for Mercury 3? Its primary use cases are any application requiring real-time AI interaction: live voice assistants, interactive tutoring and education platforms, autonomous web navigation agents, real-time search engines, and low-latency customer support chatbots.
5. How can I try Mercury 3? You can test the model for free in a chat interface at chat.inceptionlabs.ai or sign up on the Inception API platform (platform.inceptionlabs.ai) which offers 100 million free tokens to start. It is also available via OpenRouter, Baseten, and Vercel.