Inception Labs Launches Mercury 2.5, Its Strongest Diffusion LLM at 1107 Tokens/sec

On September 9, Inception Labs released Mercury 2.5, announced personally by CEO and Stanford professor Stefano Ermon. The company calls it the strongest diffusion LLM (dLLM) on the market and the largest diffusion language model known to have been trained, with a 40% intelligence gain over Mercury 2, a 260k context, and over 1107 tokens/sec on common NVIDIA GPUs.

Confirmed

• Mercury 2.5 is available via the official API, OpenRouter and Baseten, with 100 million free tokens for new accounts
• Launch promo pricing is $0.04/M input and $0.15/M output tokens (about 20% off)
• Deployment case: OpenCall AI runs voice agents on real patient calls; after switching from an AI inference chip solution to Mercury, Ermon says p99 latency dropped under 1 second and voice latency can be under 200ms

Why it matters

• The diffusion architecture dramatically raises inference speed while keeping generation quality, attacking the latency bottleneck of autoregressive LLMs
• Low latency suits delay-sensitive use cases like real-time voice agents, at competitive pricing

2026-09-09 ~ 2026-09-10 · 7 related posts

• [source] Inception ships Mercury 2.5: most capable diffusion LLM at 1,107 tokens/sec, 260K context — StefanoErmon · 2026-09-09
• [source] OpenCall runs live voice agents on Mercury 2.5, cutting p99 latency to 1 second — StefanoErmon · 2026-09-09
• [source] Mercury 2.5 diffusion LLM launches at 80% off, p99 voice latency cut from minutes to 1s — StefanoErmon · 2026-09-09
• Inception ships Mercury 2.5: diffusion LLM with 40% intelligence jump at 1,100 tokens/sec — timshi_ai · 2026-09-09
• Inception Labs ships Mercury 2.5, a diffusion LLM with 40% intelligence jump over 1,100 tokens/sec — Azaliamirh · 2026-09-09
• Inception launches Mercury 2.5: diffusion LLM with 40% intelligence jump, 1,100+ tokens/sec — volokuleshov · 2026-09-10

1 near-duplicate retellings: volokuleshov