Hacker News
Mercury 3: Diffusion at scale
AIFirst seen 212 points83 comments
Inception has released Mercury 3, a large AI language model that makes text by diffusion instead of the usual step-by-step method. The company says it is smarter than Mercury 1 and runs very fast.
What happened
Inception launched Mercury 3, which it calls the most capable diffusion-based LLM released so far. An LLM is a large language model, the kind of AI that reads and writes text. The company says Mercury 3 matches or beats two fast, lower-cost models, GPT-6 Luna and Claude Haiku 5.5, while running about 4 to 8 times faster.
Key numbers
- Speed: 1,068 tokens per second on NVIDIA GPUs. A token is a small piece of a word.
- Size: more than 40 times larger than Mercury 1 in total parameters.
- Context: 1 million tokens, up from 260K in Mercury 2.5. Context is how much text the model can keep in mind at once.
- Price: $0.20 per million input tokens and $0.70 per million output tokens. At launch, the price is half off.
- Test result: on 50 live web-navigation tasks (Mind2Web), Mercury 3 completed 50% of them.
Diffusion versus the usual method
Most AI language models write one token at a time, from left to right. Diffusion models instead start with a rough version and improve it in several passes. When Mercury 1 launched, the open question was whether this method could scale up and compete on hard tasks. Inception says Mercury 3 shows that it can.
Why it matters
Speed matters most for AI agents, which are programs that take many steps, like browsing websites. Each slow step adds up. Early partners say fast replies help with voice support, tutoring, and search. If the company's claims hold up, faster and cheaper models could make AI tools feel more responsive for everyday users.
What to watch
- Independent tests that check the company's quality and speed claims.
- Whether the discounted launch price stays after the offer ends.
- How well diffusion models handle hard tasks compared with standard models.
- Whether more companies adopt Mercury 3 in real products.
Glossary
- Inception
- The company that makes the Mercury diffusion language models.
- LLM
- Large language model: an AI system trained on large amounts of text to read and write language.
- Diffusion
- A method that starts with rough output and improves it step by step, instead of writing one token at a time.
- Mercury 3
- Inception's newest diffusion language model, released with a focus on speed and quality.
- Context
- How much text an AI model can consider at one time.