With the arrival of 'Mercury 2.5,' an ultra-high-speed AI generating 770 to 1,107 tokens per second, the response speed of real-time conversational AI services is expected to improve dramatically.
Imagine this: You’re chatting with a friend on a messenger app, and before you even finish your question, they’ve already provided the answer, and they summarize a long message you sent in just one second. It would feel as if someone who thinks at the exact same speed as you is sitting right next to you, not an AI.
Recently, this is becoming a reality in the artificial intelligence industry. The protagonist is ‘Mercury 2.5,’ a new language model released by Inception Labs. Let’s look at how this model, which answers at ‘light speed’—beyond simply being smart—will change our daily lives. Source: Inception Labs Blog
Why It Matters
The AI chatbots we use daily sometimes need ‘thinking time’ before providing an answer. Have you ever seen letters appear on the screen one by one, ‘tap, tap, tap’? This is because AI traditionally creates words sequentially.
However, Mercury 2.5 has completely flipped the script on this speed. According to measurements by AI analysis firm ‘Artificial Analysis,’ this model generates approximately 770.4 tokens (the minimum unit for AI text processing) per second. Source: Sam Codeman Developer reports suggest it can reach up to 1,107 tokens per second in specific environments. Source: Pradha This is a staggering 7 times faster than existing AI models. Source: AIPULSEN
Simply put, Mercury 2.5 finishes in just one second what would take a traditional AI seven seconds to ponder. Once this speed is realized, the subtle latency we feel when talking to AI via voice will disappear. Real-time translators or coding assistants will react instantly, just like real humans. Source: TechBriefe
The Explainer
Traditional large language models (LLMs) mainly used a ‘Transformer’ (an AI architecture that grasps relationships between words in a sentence) approach. This is similar to reading a book honestly from the first sentence to the last, so there were clear limitations to speed.
In contrast, Mercury 2.5 uses a unique technique called the ‘Diffusion’ method. Source: OpenRouter To use a metaphor: if a traditional AI is a ‘diligent writer’ who writes a novel one word at a time, Mercury 2.5 is more like a ‘skilled painter’ who sets the overall composition first and fills in the details all at once.
This technology is technically called ‘parallel token generation and refinement.’ Because it contemplates and refines multiple words simultaneously, it can produce results much faster. Source: Benchable
Where We Stand
Currently, Mercury 2.5 is not just fast; it possesses a massive context window of 260,000 tokens. Source: Build Fast with AI Blog This means it can read and remember the information contained in dozens of thick books at once. It also supports structured data output, tool use, and complex logical reasoning, making it highly practical for professional work. Source: Build Fast with AI Blog
However, technical figures may vary depending on measurement methods, and Mercury 2.5 is at a stage where it needs further user verification to see if it holds a practical performance edge over existing massive models in highly complex, specialized tasks. Source: 6ic
What’s Next
From now on, AI services will compete not just on ‘how smart’ they are, but on ‘how quickly’ they respond. Models like Mercury 2.5 are expected to shine particularly in fields where immediate reaction is crucial, such as voice assistants, real-time consultation, and NPCs (non-player characters) in games. Source: CallMissed The day is not far off when your smartphone will be equipped with an AI assistant that reacts faster, allowing you to hear an answer the moment you finish asking a question.
References
- Sam Codeman, Mercury 2.5 770 tokens/sec benchmark
- 6ic, Mercury 2.5 LLM achieves 770 tokens per second
- Pradha, Mercury 2.5: Inception’s push for high-speed diffusion LLMs
- TechBriefe, Mercury 2.5 LLM inference speed reaches 770 tokens per second
- VK.ru, Mercury 2.5 LLM reaches 770 tokens per second
- Inception Labs, Introducing Mercury 2.5
- Kunal Ganglani, Mercury 2.5 770 tokens/sec benchmark (2026)
- Build Fast with AI Blog, Mercury 2.5 AI model review
- AIPULSEN, Mercury 2.5 LLM achieves 770 tokens per second
- CallMissed, Mercury 2.5 diffusion LLM
- Benchable, Inception: Mercury 2.5 AI model details and benchmarks
- Telegram (Hacker News Feed), Mercury 2.5 LLM reaches 770 tokens per second
- OpenRouter, Mercury 2.5 API pricing and benchmarks
- Much cheaper price
- Overwhelming text generation speed
- Offline-only operation
- Transformer-only model
- Diffusion-based model
- Sequential simple arithmetic model
- 260,000 tokens
- 100,000 tokens
- 500,000 tokens