Inception
Category: Foundation Models / LLMs
Palo Alto AI lab building Mercury, the first commercially available family of diffusion large language models (dLLMs). Its models generate text in parallel and run at more than 1,000 tokens per second. Inception was founded in 2024. The company is led by Stefano Ermon. Based in Redwood City, United States. Team size: 11-50. Total funding raised: $50M. Latest round: Seed. Key investors include Menlo Ventures, Mayfield, M12 (Microsoft), NVentures (NVIDIA), Snowflake Ventures, Eric Schmidt, Patrick & John Collison, Andrej Karpathy, Nat Friedman & Daniel Gross, Andrew Ng.
- Founded
- 2024
- Headquarters
- Redwood City, United States
- Team size
- 11-50
- Total funding
- $50M
Value proposition
Diffusion LLMs refine output coarse-to-fine in parallel instead of generating one token at a time. The company claims 5-10x faster inference, sub-300ms time to first token and up to 70% lower cost per task, at quality comparable to speed-optimized frontier models. This suits latency-sensitive and agentic workloads.
Products and solutions
Mercury 3 (Oct 2026): 1M-token context, launch price $0.10 / $0.35 per 1M input/output tokens, Mercury 2.5, Mercury 2 (reasoning dLLM, Feb 2026), Mercury and Mercury Coder (2025), Mercury Chat, Inception API Platform, enterprise deployments.
Unique value
The first commercial-scale diffusion LLM family, from the researchers who helped invent diffusion models. It offers frontier-class speed-optimized quality at more than 1,000 tokens/s on commercial NVIDIA GPUs.
Target customer
Developers and enterprises building latency-sensitive AI applications, such as coding assistants, agents, voice and real-time apps, through an API and cloud marketplaces.
Industries served
Software development / coding tools, AI agents, enterprise applications, voice and real-time AI, cloud / AI platforms
Technology advantage
Founders' research lineage is cited on the company's research page: diffusion models, Flash Attention, DPO, Decision Transformers, masked/block/remasking diffusion and d1 reasoning. Parallel coarse-to-fine generation gives a different speed-cost curve than autoregressive LLMs. The models run on standard NVIDIA GPUs without custom hardware.
How they differentiate
Uses diffusion rather than autoregressive generation. Mercury 3 claims to beat GPT-6 Luna, Claude Haiku 5.5 and GLM-5.3 Flash on agentic benchmarks while running above 1,000 tokens/s. Pricing is lower than frontier speed-tier models. These benchmarks are self-reported.
Main competitors
Google (Gemini Diffusion / Gemini Flash), OpenAI (speed-optimized GPT models), Anthropic (Claude Haiku)
Key partnerships
Microsoft Azure AI Foundry (Mercury available), NVIDIA (NVentures investor, runs on NVIDIA GPUs), Snowflake (Snowflake Ventures investor), third-party inference platform integrations
Notable customers
Augment Code
Major milestones
2024: founded by Stanford, UCLA and Cornell professors. Feb 2025: emerged from stealth with Mercury, the first commercial dLLM. Mercury became available on Azure AI Foundry. Nov 2025: raised $50M seed led by Menlo Ventures. Feb 2026: Mercury 2 reasoning dLLM. Mercury 2.5 followed. Oct 2026: launched Mercury 3 with a 1M-token context window and 50% launch discount.
Growth metrics
Mercury 3 is reported to be more than 40x larger than earlier Mercury models and runs at about 1,068 tokens/s. The release cadence went from Mercury (Feb 2025) to Mercury 2 (Feb 2026), Mercury 2.5, then Mercury 3 (Oct 2026).
Market positioning
Pioneer and leading commercial vendor of diffusion LLMs, positioned as the speed and cost frontier for production and agentic AI.
Geographic focus
Global (US-based; API available worldwide and via Azure)
About Stefano Ermon
Associate Professor of Computer Science, Stanford University, where his lab did foundational work on diffusion models; Cornell University (PhD). Co-inventor of key diffusion and generative-modeling techniques; CEO of Inception since Jul 2024.
Latest news about Inception
More Foundation Models / LLMs companies
Official website: https://www.inceptionlabs.ai