Revising text during generation
Inception, recorded in the timeline as Inception Labs, is a research and product company developing diffusion language models. Its Mercury family uses a different generation process from the familiar left-to-right sequence of an autoregressive language model: it refines an output over successive steps, changing multiple token positions in parallel. That approach makes generation speed a central part of the company's product and research work. [1][2]
The distinction concerns how an answer is produced. Inception still offers the resulting models for familiar applications such as coding, retrieval-grounded answers and tool use. Its original announcement described Mercury as a replacement within those workflows, with the diffusion process taking place inside the model rather than requiring the user to build a new kind of application. [2]
Mercury began with code
Mercury's public debut on February 26, 2025 included Mercury Coder, a code-generation model available to try in a playground. Inception's announcement, written by CEO Stefano Ermon, also offered enterprise access through APIs and on-premise deployments. It reported generation above 1,000 tokens per second on NVIDIA H100 GPUs, alongside coding benchmark results. [3][2]
The release made diffusion language generation a commercial offering readers could examine beyond a paper or demonstration. Its benchmark table also matters more than a speed headline alone: generation rate and answer quality are separate measurements. The comparisons in the announcement were Inception's own evidence and do not establish that the model was faster or more accurate on every customer's workload. [2]
Scaling the same approach
Mercury 3 arrived on October 19, 2026 with a one-million-token context window, up from the 260,000 tokens cited for Mercury 2.5. Inception made it available on its own API platform and through OpenRouter, Baseten and Vercel. The release reported 1,068 generated tokens per second on NVIDIA GPUs and described uses in browser agents, tutoring, search and customer support. [4]
Those applications share a reason to care about latency: a user or another step in a workflow is waiting for the output. Inception's release supports its positioning around that constraint, while its selected customer accounts and benchmark comparisons remain vendor-published evidence. Throughput, response time and the reliability of a complete agent are different properties; a faster generator alone does not establish the last of them. [4]
Sources
- About
Inception · retrieved Oct 23, 2026
- Introducing Mercury, the World’s First Commercial-Scale Diffusion Large Language Model.
Inception · retrieved Oct 23, 2026
- Inception emerges from stealth with a new type of AI model
TechCrunch · Feb 26, 2025
- Diffusion at Scale: Introducing Mercury 3
Inception · Oct 19, 2026