Inception's upcoming next-gen AI, 'Mercury 3,' uses diffusion model-based technology to grasp entire contexts at once, expected to match the performance of existing models.
Imagine you need to write a long and complex letter to a friend. Until now, AI has functioned as if it were writing non-stop, one character at a time, worrying only about the next word. If it made a mistake, it was difficult to go back and fix it. But what if we could envision the entire content of the letter in our heads at once, and refine it while observing the overall context? Inception’s upcoming next-gen AI model, ‘Mercury 3,’ is heralding just such an era. Source: Batalla por Inception: Microsoft vs SpaceX en la carrera de IA
Why is this important?
Most AI we use today are ‘autoregressive’ models. Simply put, they predict the next word based on the word just written. However, this method has the drawback of easily forgetting the intention of the beginning if the sentence gets long, or committing logical errors.
On the other hand, Large Language Models (dLLMs) based on ‘diffusion models,’ such as Mercury 3, view the entire sentence at once. This is very similar to how we conceive the overall logical structure first and then fill in the details when we write. This means AI will be able to ‘think’ and write much more intelligently. Source: DiffuCoder: Understanding and Improving Masked Diffusion Models…
Easy to understand: AI resembling a photo editing app
To understand diffusion models more easily, think of the ‘noise reduction filter’ in a photo editing app. It is similar to the process of sprinkling a lot of messy noise on a photo to make its form unrecognizable, and then gradually removing the noise to find the original clear photo again. Source: DiffuCoder: Understanding and Improving Masked Diffusion Models…
Language models are the same. AI starts with a jumble of words (a state mixed with noise). Then, through repetitive processes, it refines those words to fit the context (removing the noise) to complete a perfect sentence. In this process, the AI does not just see a part of the sentence but deals with the ‘entire sequence’ at once. Through this, it does not miss the overall logical flow and can continuously improve the content (iterative refinement) if necessary. Its true value is revealed especially in coding or long-form writing that requires complex logic. Source: DiffuCoder: Understanding and Improving Masked Diffusion Models…
Current status: Evolving technology
In September 2026, Inception introduced the diffusion model-based ‘Mercury 2.5,’ proving that this method is sufficiently feasible. Source: Mercury2.5 API Pricing 2026 — Cost per 1M Tokens · AI//COST And now, it is poised to announce ‘Mercury 3,’ which has evolved further. According to Inception, this Mercury 3 is expected to provide quality comparable to the flagship autoregressive models currently considered the industry standard. Source: Batalla por Inception: Microsoft vs SpaceX en la carrera de IA
What will happen in the future?
Once Mercury 3 is officially released, the way we communicate with AI will change significantly. Especially in environments requiring complex calculations, the AI’s response speed and accuracy are expected to improve drastically. As AI moves beyond simply ‘spitting out’ words one by one to ‘understanding’ the entire context and providing logical conclusions, its role as a professional partner beyond a simple conversational partner will grow even larger. Source: Batalla por Inception: Microsoft vs SpaceX en la carrera de IA
MindTickleBytes’ AI Reporter Perspective
This change introduced by Inception signifies a transition in the fundamental approach of how AI handles language, going beyond mere performance improvement. When an AI that used to move ‘one word at a time’ begins to ‘view the whole,’ AI will be one step closer to becoming an intellectual colleague that worries and thinks with us, no longer just a simple statistical probability calculator.
References
- Mercury2.5 API Pricing 2026 — Cost per 1M Tokens · AI//COST (https://aicost.tools/llm-cost/inception/mercury-2-5/)
- DiffuCoder: Understanding and Improving Masked Diffusion Models… (https://arxiv.org/pdf/2506.20639)
- Batalla por Inception: Microsoft vs SpaceX en la carrera de IA (https://sudonull.com/es/battle-for-inception-microsoft-vs-spacex-in-the-ai-race)
- They only generate words sequentially
- They enable global planning by grasping and revising context at once
- They are unconditionally faster than existing models
- Much lower level than existing models
- Comparable to flagship models
- Still in the early stages of development
- Whether they generate words one by one sequentially or view the whole
- The difference in language used
- The amount of training data