Overview
Mercury 3 is a diffusion-based large language model released by Inception Labs on October 19, 2026. The model uses a diffusion architecture, differing from token-by-token autoregressive generation. The company claims a generation speed exceeding 1000 tokens/s, a context window of 1M tokens, and a model size more than 40 times that of Mercury 1. The product focuses on low latency and high throughput, targeting application scenarios with requirements for response speed and concurrent processing. Inception Labs provides this model service through its API platform. The official website is inceptionlabs.ai, and specific pricing is subject to the official website.
Key Features
- Diffusion-based generation architecture: Uses a diffusion architecture rather than token-by-token autoregression. The company states that this design helps reduce generation latency and improve throughput performance.
- Thousand-level token generation speed: The company claims a generation speed exceeding 1000 tokens/s, suitable for interactive scenarios with high requirements for real-time output.
- Million-level context window: The context window is 1M tokens, capable of accommodating longer documents and multi-turn conversation content, reducing the need for segmented processing.
- Significant expansion of model scale: The model size is more than 40 times that of Mercury 1, which the company uses as the baseline specification for capability improvement.
- API platform access: Inception Labs provides the Mercury 3 service through its API platform, and developers can integrate and call it according to the official website guidance.
Use Cases
- Call the Mercury 3 API in real-time conversation applications, using its low-latency characteristics to shorten the time users wait for replies.
- When handling long document summarization or analysis tasks, use the 1M token context window to load complete materials at once.
- For high-concurrency online service scenarios, call the model in batches through the API to support processing of larger request volumes.
- Integrate the model into code assistance or text generation tools, relying on the thousand-level tokens/s generation speed to improve output efficiency.
- Build retrieval-augmented generation workflows that require fast responses, placing retrieved content and questions together into the long context for processing.
Pros
- The company claims a generation speed exceeding 1000 tokens/s, with clear specifications for low-latency output.
- The context window reaches 1M tokens, suitable for tasks that need to process long texts or large amounts of reference materials.
- Uses a diffusion architecture, offering a different technical path choice from token-by-token autoregressive generation.
- The model size is more than 40 times that of Mercury 1, which the company uses to reflect the scale of generational specification expansion.
- Inception Labs provides an API platform, and developers can learn about access methods and billing information through the official website.
Pricing
Mercury 3 is available through the Inception Labs API platform and billed by usage. See the official site for current pricing, quotas and access details.
Summary
Mercury 3 is a diffusion-based large language model launched by Inception Labs. The company claims a generation speed exceeding 1000 tokens/s, a context window of 1M tokens, and a model size more than 40 times that of Mercury 1. Its diffusion architecture and low-latency, high-throughput positioning make it suitable for scenarios such as real-time interaction, long document processing, and high-concurrency API calls. Specific pricing and access details are subject to the official website.
Version History
- Mercury 3 launches: diffusion LLM claiming over 1,000 tokens/s and 1M context (2026-10-19): Inception Labs released Mercury 3, a diffusion-based large language model, on October 19, 2026. The company claims generation speeds above 1,000 tokens per second, a 1M token context window, and a model scale more than 40 times that of Mercury 1. Instead of token-by-token autoregressive decoding, the diffusion architecture produces text in parallel, targeting low-latency and high-throughput workloads. It is available through the Inception Labs API platform.