Mercury AI is not one product: the name is used by Mercury AI GmbH, a small chatbot and voicebot vendor in Bielefeld, Germany, while Mercury is also a family of diffusion language models from the U.S. company Inception, first shown in February 2025. No source checked for this article links the two organizations.
Key takeaways
- Mercury AI GmbH is a conversational AI vendor based in Bielefeld, Germany, that describes itself as a spin-off from CITEC at Bielefeld University and as on the market since 2016.
- Inception's Mercury models generate text by diffusion, refining many tokens in parallel instead of writing one token at a time.
- The Mercury technical report, submitted to arXiv on June 17, 2025, lists speeds of 1,109 and 737 tokens per second for its two Coder models on NVIDIA H100 GPUs, measured by Artificial Analysis.
- Inception announced a USD 50 million funding round on November 6, 2025, led by Menlo Ventures.
- Most capability claims for both products come from the companies themselves.
Mercury AI GmbH: a German conversational AI vendor
The company's own documentation describes Mercury AI GmbH as a European platform for controlled, process-integrated conversational AI. It is an owner-managed firm with 15 people, registered in Bielefeld, and states that it came out of CITEC, the Cluster of Excellence Cognitive Interaction Technology at Bielefeld University. The product has four parts: a chat channel, a voice channel for telephone systems, a handover to human agents in live chat, and analytics. A no-code studio lets business teams design dialogues, and connectors link the platform to ERP, CRM, shop, and ticketing systems.
The company says answers are generated only from sources the customer has approved, that data is hosted in Germany on AWS Frankfurt without transfer to third countries, and that customer data is not used for training. On certification, the company says its data centers are certified to ISO 27001 while it works in line with the standard, and that Mercury AI GmbH itself does not hold the certificate. These are vendor statements, and no independent audit was reviewed for this article.
Mercury from Inception: a different kind of language model
Inception, a Stanford spinout led by professor Stefano Ermon as co-founder and chief executive, introduced Mercury Coder in February 2025. Trade coverage on February 26, 2025 described it as the first commercial-scale diffusion language model. Diffusion is the method behind many image generators: output begins as noise and is refined over several passes. Inception describes the text version as coarse to fine, with the model able to change several tokens at once and to correct earlier mistakes, where a conventional autoregressive model commits to each token in left-to-right order.
The technical report, "Mercury: Ultra-Fast Language Models Based on Diffusion," was submitted to arXiv on June 17, 2025. It describes Transformer models trained to predict multiple tokens in parallel and focuses on Mercury Coder in Mini and Small sizes. The abstract reports, from Artificial Analysis tests on H100 GPUs, 1,109 tokens per second for Mini and 737 for Small, and says the models match speed-optimized frontier models on quality while running up to 10 times faster. It also reports a second-place quality ranking and the fastest speed on Copilot Arena. The abstract does not name the individual coding benchmarks.
Side-by-side summary
| Item | Mercury AI GmbH | Inception Mercury |
|---|---|---|
| What it is | Chatbot and voicebot platform for businesses | Family of diffusion large language models |
| Base | Bielefeld, Germany | United States |
| Origin | Spin-off from CITEC, Bielefeld University | Research by Stefano Ermon and colleagues |
| Public technical record | Company documentation | arXiv report, June 17, 2025 |
| Typical customer | Companies running customer service channels | Developers who call the models through an API |
| Claims to verify independently | Hosting, certification, automation rates | Speed and quality comparisons |
Funding and later versions
On November 6, 2025, press coverage reported that Inception raised USD 50 million in a round led by Menlo Ventures, with Mayfield, Innovation Endeavors, NVentures, M12, Snowflake Ventures, and Databricks Investment taking part. Later versions of Mercury have been announced since then, including reasoning-capable ones, but their figures appear mainly on the company's own pages without published evaluation methods, so they are not repeated here. Pricing and context-window sizes also change from release to release and should be read from the provider's current documentation.
How to read vendor claims about either product
Three questions apply to both. First, who measured it: an independent evaluator such as Artificial Analysis gives a stronger basis than a company blog, though the Mercury report itself cites the evaluator's numbers. Second, on what hardware and workload: a speed figure on H100 GPUs for code completion does not carry over to a customer-service dialogue on other hardware. Third, what is held constant: a model that is 10 times faster at equal quality on a benchmark may differ on a specific task, so a small test on real inputs is more informative than a headline ratio. For the chatbot vendor, the equivalent question is how often the system hands a conversation to a human and how that was counted.
FAQ
Is Mercury AI GmbH the same as Inception's Mercury?
No source reviewed links them. Mercury AI GmbH sells conversational AI software from Germany, while Mercury is a diffusion language model family from Inception in the United States.
What does a diffusion language model do differently?
It starts from a rough version of the whole output and refines it over several passes, changing many tokens in parallel. Most other language models produce one token after another.
Are the speed numbers independent?
The 1,109 and 737 tokens-per-second figures in the technical report are attributed to Artificial Analysis evaluations on H100 GPUs. Results on other hardware or tasks were not checked.
Where is customer data stored for Mercury AI?
The company states hosting exclusively in Germany on AWS Frankfurt. That is a vendor statement, and contract terms should be checked directly.
Related reading
- Who is behind artificial intelligence
- AI algorithms for the traveling salesman problem
- The origins of artificial intelligence
- Whether AI can experience emotions
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Mercury AI llms.txt, Mercury AI GmbH, 2026
- Mercury: Ultra-Fast Language Models Based on Diffusion, Inception Labs, arXiv:2506.17298, 2025
- Inception Labs launches Mercury, the first commercial diffusion-based language model, Maginative, February 26, 2025
- Inception raises USD 50 million, SiliconANGLE, November 6, 2025
- Introducing Mercury, the first commercial-scale diffusion large language model, Inception, 2025