Inception Launches Mercury 3, a Diffusion LLM Claiming 1,000+ Tokens/sec
sedielem · x · 2026-10-20
Inception, founded by Stanford professor Stefano Ermon, released Mercury 3, calling it the most capable diffusion LLM to date.
- 3x more intelligent than Mercury 1 on standard benchmarks, ahead of GPT-6 Luna and Claude Haiku 5.5
- Runs at over 1,000 tokens/sec — same speed as Mercury 1 at more than 40x the size, arguing diffusion keeps its speed edge as it scales
- Early partners are already running it in browser agents and real-time tutoring
More from Models
- Codex reset perks 'not free': user claims they just push back your usual reset — rudrank · 2026-10-20
- Speculative decoding speeds up dense models more than MoEs, as intuition suggests — mervenoyann · 2026-10-20
- New decision model basal-omni-1.0-ultra claims near-frontier intelligence at much lower cost — tristanbob · 2026-10-20
- Inception Labs launches Mercury 3, a diffusion LLM claiming 1000 tok/s inference — cmrdporcupine · 2026-10-20
- Reasoning models can steer their chain of thought far better than earlier tests suggested, paper finds — dair_ai · 2026-10-20
- Same sources, same question: how Claude, NotebookLM and ChatGPT differ on one knowledge base — kbavandi · 2026-10-20