Mercury 3: Diffusion at scale
- ID
- 37274
- Status
- summarized
- Published
- 20 Oct 2026, 1:27 AM
- Fetched
- 20 Oct 2026, 11:06 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://www.inceptionlabs.ai/blog/introducing-mercury-3
- Source URL
- https://hnrss.org/best
Summary
- Score
- 6.5
- Created
- 20 Oct 2026, 11:06 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_userssaas_startup_founders
What happened
Inception launched Mercury 3, a diffusion-based LLM it says beats GPT-6 Luna and Claude Haiku 5.5 while running 4–8x faster at 1,068 tokens/sec, with 1M context and $0.20/$0.70 per 1M input/output tokens (50% off at launch). The company says it is over 40x larger than Mercury 1 and cites early partners using it for browser agents, tutoring, support, and data engineering; a production web-agent test completed 50% of 50 live Mind2Web tasks versus GPT-6.1-Sol. The HN thread has 226 points and 85 comments.
Why it matters
If you run latency-sensitive agent loops or high-volume chat/tutoring, Mercury 3 is a concrete option to benchmark against GPT-6 Luna/Claude Haiku 5.5: 1,068 tok/s, 1M context, and launch pricing of $0.10/$0.35 per 1M input/output tokens. Do not migrate on Inception's vendor-run Mind2Web result alone; rerun a comparable 50-task navigation eval and compare token volume and per-task cost at your latency budget. There is no Malaysia/SEA-specific detail in the post.
Discussion angle
Diffusion vs autoregressive for agent loops: would 1,068 tok/s and lower per-token cost justify testing a less-proven model family, and what eval would you require before switching?