Inception introduced Mercury 2, described as the world's fastest reasoning language model, built on diffusion-based parallel refinement rather than autoregressive sequential decoding. The post claims 5x faster generation, 1,009 tokens/sec on NVIDIA Blackwell GPUs, a 128K context window, native tool use, tunable reasoni The launch post argues that diffusion-based reasoning enables reasoning-grade quality inside real-time latency budgets, because higher intelligence no longer requires linearly more sequential test-time compute. An included quote from NVIDIA's Shruti Koparkar highlights Mercury 2 surpassing 1,000 tokens/sec on NVIDIA GP

Inception announced a $50M financing round led by Menlo Ventures, with participation from Andrew Ng, Andrej Karpathy, and other top venture funds and angels. The post also introduced a refreshed Mercury diffusion-based LLM (dLLM) delivering improvements in coding, instruction following, mathematical problem solving, an The post explains the technical premise of Mercury dLLMs: unlike autoregressive LLMs that generate tokens sequentially, diffusion-based models generate entire blocks of language in parallel via an iterative denoising process, yielding up to 10x faster and more efficient inference while maintaining best-in-class quality

The official Inception blog index lists a series of 2026 Mercury 2 product milestones, including Mercury 2 for Search (Aug 11, 2026), a throughput and builder update (Jul 29, 2026), a Mercury 2 reasoning launch (Jul 14, 2026), Mercury 2 availability on Azure AI Foundry (Jun 24, 2026), Mercury 2 powering real-time subag Distribution and partnership highlights from the index include Mercury 2 launching on Azure AI Foundry to bring dLLM inference to enterprise developers, and a SearchBlox integration that delivers sub-second GenAI responses for enterprise search workloads. The index also references a Mercury diffusion LLM availability m