Inception

The three founders had developed a teaching and research relationship at Stanford before holding academic positions at Stanford, UCLA and Cornell. The company dates its founding to 2024. Mayfield says it led an investment in July 2024. The team sought to take text-diffusion research into commercial services requiring substantial computing resources, capital and engineering.⁠[1][2][3]

Inception publicly introduced the Mercury diffusion language-model family, initially targeting programming with Mercury Coder Mini and Small. It opened an online playground and partnered with LambdaLambdaLambda is an American AI cloud computing company founded in 2012 by Stephen Balaban and Michael Balaban. It initially sold GPU workstations and servers for deep learning, later shifted to renting out cloud computing capacity based on Nvidia GPUs, and builds supercomputers and “AI factories” for AI training and inference.Open the full entry for computing infrastructure, alongside plans for enterprise APIs, private deployments and customized models. Its approach generates and repeatedly refines text in parallel to reduce the waiting inherent in token-by-token generation.⁠[4]

In a November 7 model-update announcement, Inception confirmed a recent $50 million financing led by Menlo Ventures, with support from Andrew Ng, Andrej Karpathy and others. The update expanded training data, improved the denoiser and training approach, and optimized inference serving. Developers could access Mercury through Inception’s API and distribution channels including OpenRouterOpenRouterOpenRouter is an AI model marketplace and gateway launched in 2023. Its API and chat interface connect models and inference providers, with routing, observability, cost management and enterprise controls. Its co-founders include Alex Atallah, Louis Vichy and Chris Clark. Stripe announced an agreement to acquire OpenRouter in August 2026.Open the full entry.⁠[5]

Mercury became available in Azure AI Foundry for enterprise deployment, with tool calling, structured output and an OpenAI-compatible interface. Organizations could run the model in their Azure environment. The announcement treated model licensing and cloud-compute charges separately.⁠[6]

Inception released Mercury 2 for reasoning workflows, with adjustable reasoning effort, a 128,000-token context window, tool calling and JSON output. It positioned the model for multi-step agents, retrieval and interactive applications, citing evaluation feedback from organizations including Wispr FlowWispr FlowWispr Flow is an AI voice-input product and brand developed by Wispr. It turns natural speech into text within applications and has expanded into meeting notes. Founded in 2021 by Tanay Kothari and Sahaj Garg and based in San Francisco, Wispr initially explored wearable voice interfaces before focusing on Flow software in 2024.Open the full entry. Speed and quality figures were reported by the company and its cited evaluators, rather than guarantees across every deployment.⁠[7]

Mercury Edit 2 targets next-edit prediction, inferring a developer’s likely subsequent changes from recent edits and surrounding context. Inception described preference-optimization training designed to align suggestions with edits developers accept, and announced availability in Zed and ProxyAI. The product focuses on the editing workflow rather than general conversation.⁠[8]

Mercury 2.5 expanded the context window to 260,000 tokens and supported parallel tool calls, structured output and adjustable reasoning effort. Inception emphasized long-context work, multi-step tasks and agent tool selection, citing Augment Code’s use in context compaction and tool retrieval. The announcement also previewed Mercury Voice and Mercury Router; those previews were distinct from general availability.⁠[9]

Mercury Voice entered general availability for real-time voice agents. It fills the language-model stage of a voice pipeline rather than replacing speech recognition and synthesis. It offers a 128,000-token context window, three reasoning-effort levels and a compatible API. Inception reported latency to the first answer token on customer prompts, a metric distinct from the end-to-end response time of a complete voice system.⁠[10]

OpenRouter lists September 30 as Mercury Decide’s release date. The product selects among explicit choices, scores inputs or makes yes/no judgments, returning structured decisions and probability information for routing, filtering and agent control. On OpenRouter it uses the System One Decisions API rather than the standard chat-completions interface. Availability and terms of the free variant depend on the platform’s current listing.⁠[11][12]

Inception applies diffusion to text generation by starting with a coarse candidate and refining multiple positions in parallel. It uses this approach to explore combinations of inference speed, cost and output quality. Distribution includes its own API, OpenRouter, Baseten and enterprise cloud services. Features, interfaces and deployment options differ by model, so one version’s performance is not a fixed property of the entire family.⁠[13][12]

Website: https://www.inceptionlabs.ai/ (opens in a new window). X: https://x.com/_inception_ai (opens in a new window). LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry: https://www.linkedin.com/company/inception-labs-ai/ (opens in a new window).⁠[12][14][1]