# Inception ships Mercury 2.5 at 1,107 tokens per second, by its count > Source: > Published: 2026-09-08 23:48:28+00:00 # Inception ships Mercury 2.5 at 1,107 tokens per second, by its count **Three university professors are betting parallel generation can cut the time and bill behind coding, search and voice agents.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [Inception Labs on X](https://x.com/_inception_ai/status/2097365772289151417) ## Why it matters Mercury 2.5 tests whether diffusion can become a practical serving architecture for high-volume agents, where latency and cost compound across every model call. [Stefano Ermon (@StefanoErmon)](https://x.com/StefanoErmon?ref=runtimewire), [Aditya Grover (@adityagrover_)](https://x.com/adityagrover_?ref=runtimewire) and [Volodymyr Kuleshov (@volokuleshov)](https://x.com/volokuleshov?ref=runtimewire) launched Mercury 2.5 on September 8th, pushing their diffusion language model research into a production-oriented API that Inception says can generate 1,107 tokens per second on widely available Nvidia GPUs. Inception detailed the release in [a launch post](https://www.inceptionlabs.ai/blog/introducing-mercury-2-5?ref=runtimewire) after announcing it [on X](https://x.com/_inception_ai/status/2097365772289151417?ref=runtimewire). Mercury 2.5 has a 260,000-token context window, adjustable reasoning, parallel tool calls and schema-aligned JSON output. List pricing is $0.20 per million input tokens and $0.75 per million output tokens, with an 80% launch discount that lowers those rates to $0.04 and $0.15. Inception has not specified when the discount ends. The release carries a larger technical bet from three academics who worked together through Stanford for nearly a decade: useful language models do not have to write like typewriters, committing to one token before producing the next. Diffusion models begin with a rough output and refine many positions in parallel, trading the rigid left-to-right sequence for repeated denoising steps. Inception says that design can lower latency and serving costs when an application makes repeated model calls. Mercury 2.5 is aimed squarely at those workloads: coding subagents, search pipelines, voice systems and routers that may invoke a model dozens of times before returning one answer. ### Three professors take diffusion out of the lab Ermon, Inception's CEO and a Stanford computer science professor, began exploring diffusion for text in his university lab. He founded Inception in summer 2024 after deciding that proving the architecture at commercial scale required more compute, engineering and capital than an academic group could supply. Grover, Inception's CTO, is an assistant professor at UCLA whose work covers generative models, reinforcement learning and scientific discovery. He completed his Ph.D. at Stanford, trained as a postdoctoral researcher at UC Berkeley and previously worked with Meta's Fundamental AI Research group. His earlier research stops included Google DeepMind, Microsoft Research and OpenAI. Kuleshov is a Cornell and Cornell Tech professor who completed his Ph.D. at Stanford under Ermon. His group has published work on masked diffusion and block diffusion, two approaches for generating and refining text without relying exclusively on conventional token-by-token decoding. He also co-founded Afresh, which applies machine learning to grocery inventory and food waste. The founders' research relationship matters because Mercury is an attempt to commercialize an architectural change, rather than wrap another provider's model in application software. Menlo Ventures described the trio as having spent nearly a decade working together before forming Inception. The three retained university roles while building a model developer that has to compete for the same scarce researchers, Nvidia capacity and enterprise workloads pursued by much larger labs. In November 2025, Inception raised a $50 million seed round led by [Menlo Ventures](https://menlovc.com/perspective/from-the-lab-to-the-frontier-the-story-behind-inception/?ref=runtimewire), according to [TechCrunch](https://techcrunch.com/2025/11/06/inception-raises-50-million-to-build-diffusion-models-for-code-and-text/?ref=runtimewire). Mayfield, Innovation Endeavors, Microsoft's M12, Snowflake Ventures, Databricks Ventures and Nvidia's NVentures participated, while Andrew Ng and Andrej Karpathy invested as angels. That investor list gives Inception capital plus potential distribution and infrastructure relationships, although no valuation was reported. Kuleshov recently laid out the technical lineage behind this work in [a diffusion LLM guide published by RuntimeWire](/article/volodymyr-kuleshov-diffusion-llm-builders-manual), covering masked denoising, variable-length generation and the quality-speed trade-off. Mercury 2.5 is the commercial test of whether those techniques survive contact with production traffic. ### The speed claim needs a denominator Inception reports 1,107 tokens per second and a 40% increase in "intelligence" from [Mercury 2](/models/inception/mercury-2). The launch material does not attach that 40% figure to a clearly named independent benchmark in its prose, and the throughput result lacks enough disclosed detail to compare it cleanly with rival claims. Batch size, output length, hardware configuration, reasoning-token accounting and network overhead can all change a tokens-per-second result. That caveat matters in a category where several developers publish four-digit speeds under different conditions. ByteDance said its Seed Diffusion Preview reached 2,146 tokens per second in 2025. Google introduced Gemini Diffusion that year as an experimental research demo rather than a broadly available production API. Celeris reported 1,664 tokens per second for Celeris-1 in July 2026 using its own workload and methodology. Those figures establish that Inception is no longer alone in pursuing diffusion for language. They do not create a reliable leaderboard. Celeris, for example, measured the older Mercury 2 at a median 324 tokens per second on its test, far below Inception's advertised rate for that generation. Celeris disclosed that the comparison mixed server-reported processing time and end-to-end timing, illustrating why headline speeds require the workload and measurement method beside them. Mercury 2.5 therefore has to win on completed tasks, predictable latency and output quality. Inception says the prior Mercury 2 accumulated thousands of developers, dozens of enterprise deployments and more than tenfold usage growth. Those remain Inception's figures. The named production examples are more useful: OpenCall uses Mercury for AI phone agents, while Augment Code uses it for context compaction, model routing and tool search. [Inception says](https://www.inceptionlabs.ai/blog/introducing-mercury-2-5?ref=runtimewire) Augment Code cut compaction latency from roughly 150 seconds to 27 seconds and reduced the associated model cost by 90%. [OpenCall says](https://www.inceptionlabs.ai/blog/introducing-mercury-2-5?ref=runtimewire) Mercury brought median response latency close to 170 milliseconds on its workload. Both are customer-reported results tied to particular implementations, rather than general performance guarantees. ### The business case is repeated inference Mercury 2.5 is available through the [Inception API](https://platform.inceptionlabs.ai/?ref=runtimewire), Baseten and OpenRouter, while the current [model catalog](https://www.inceptionlabs.ai/models?ref=runtimewire) labels Mercury 2.5 as a preview. Inception is offering 100 million free API tokens and enterprise configurations including dedicated capacity, autoscaling, configurable data retention and compliance controls. The aggressive launch pricing is aimed at removing friction from workload tests. An 80% discount makes it inexpensive for engineering teams to run their own prompts, although the eventual economics depend on how long that discount remains and whether Mercury can reduce the total number of seconds and GPU cycles required per completed task. That is where Ermon, Grover and Kuleshov have placed their commercial bet. Agent systems magnify small delays because each planning, retrieval, routing and summarization step waits on another inference call. A model that saves hundreds of milliseconds on each call can change how much work fits inside an interactive session. A fast model that produces unreliable tool calls simply moves the failure earlier. Inception is already previewing Mercury Voice, with a stated target below 170 milliseconds to the first token, and Mercury Router, which would select among open and closed models based on quality, speed and cost. Inception has not set general availability dates for either product. Mercury 2.5 gives the founders a broader, cheaper model to put in front of production developers while diffusion language models still have a narrow window to establish themselves as an architecture rather than a benchmark curiosity. The decisive numbers will come from customer workloads: full-task latency, error rates and cost per successful outcome. Tokens per second gets Mercury 2.5 into the test. It does not decide whether the model stays there.