Vibeleaderboard
Inception

ai lab

Inception

Inception matters because it is testing a genuinely different route to production language models. Mercury shows that diffusion decoding can move the speed frontier on standard NVIDIA H100 hardware, while Mercury 2 exposes the remaining tradeoff: extremely high output throughput does not by itself produce frontier-level intelligence or instant first-token response.5,2,1

Official site ↗Models in the Atlas →

Profile

Overview

A company built around diffusion for text

Inception is a Palo Alto AI research and product company founded in 2024 by Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov. The three academics had worked together for years across Stanford, UCLA, and Cornell on generative modeling, representation learning, and machine learning systems. They formed the company around a specific architectural bet: text models could generate through iterative denoising instead of committing to one token at a time.1,9

A research lineage, not a solitary invention

That bet grew from a longer research lineage rather than a new label applied to a standard language model. Ermon and collaborators developed score-based generative modeling through stochastic differential equations before Inception existed. Later Block Diffusion research, co-authored by Ermon and Kuleshov with other researchers, explored a hybrid between autoregressive and diffusion language models. Inception draws on that lineage, but the company should not receive sole credit for the wider field of diffusion modeling.8,7,6

Mercury becomes a commercial model family

Mercury was Inception's first commercial model family. Mercury Coder applied parallel token prediction to software tasks, followed by a general chat model and API. The associated technical report describes transformer-based diffusion models trained to predict multiple tokens in parallel. Inception raised a $50 million seed round in 2025 to develop the approach, and early integrations put Mercury into several coding tools.5,1

Mercury 2 clarifies the tradeoff

Mercury 2 extended the family into reasoning and broader production workloads. Artificial Analysis measured 928.2 output tokens per second, compared with a 100.0 token-per-second median for similarly priced reasoning models, but assigned an Intelligence Index score of 22 and measured a 3.23-second time to first token against a 2.05-second median. That combination defines the company more accurately than a claim of general frontier leadership: Inception has demonstrated exceptional decode throughput, while capability breadth and ecosystem adoption remain developing evidence.2,4

Notable contributions

  1. 01A commercial diffusion-language-model APIMercury turned parallel denoising for text into hosted code and chat APIs. The claim is limited to Inception's commercial model family, not the invention of diffusion or masked language modeling.5,1
  2. 02Third-party measurement of diffusion code modelsArtificial Analysis measured Mercury Coder Mini above 1,000 output tokens per second and reported Coding Index results for both Coder models. Paired with public API access, those measurements gave developers a concrete test of diffusion generation in latency-sensitive software workflows.3,5,1
  3. 03Reasoning through parallel denoisingMercury 2 applies diffusion generation to a reasoning model and exposes the result through a production API. Independent measurements confirm its exceptional output speed while also showing that intelligence and first-token latency remain separate constraints.2,4