Inception Mercury 3 Reaches 1,000‑tok/s Diffusion on LLM
claude
| Source: HN | Original article
Inception released Mercury 3, a diffusion LLM processing over 1,000 tokens per second and claimed to outperform GPT‑6 Luna and Claude Haiku 5.5 while running 4–8× faster.
Inception Labs unveiled Mercury 3, a diffusion‑based large language model that it claims shatters the speed barrier for generative AI. The company announced the model on October 19, noting that Mercury 3 can generate more than 1,000 tokens per second—four to eight times faster than leading autoregressive systems such as GPT‑6 Luna and Claude Haiku 5.5, while also delivering higher benchmark scores.
The launch marks the latest step in Inception’s push to replace the traditional “one‑token‑at‑a‑time” approach with diffusion, a technique that predicts multiple tokens in parallel. According to Inception, this architecture preserves quality even as model size scales, allowing Mercury 3 to retain the speed advantage of earlier, smaller diffusion models while achieving a new level of capability. CEO and co‑founder Stefano Ermon highlighted the model’s suitability for real‑time use cases, citing early partners that are already integrating it into browser agents and interactive tutoring tools.
Why it matters is twofold. First, the speed gains could reshape workloads that depend on low‑latency text generation, from conversational assistants to on‑device applications. Second, the performance edge over established models signals that diffusion may become a viable alternative to the autoregressive paradigm that has dominated the field for years, potentially prompting other labs to explore similar designs.
The next weeks will reveal whether Mercury 3’s claims hold up under independent testing and how quickly developers adopt the diffusion workflow. Industry observers will be watching benchmark releases, pricing strategies, and any follow‑up models that Inception may roll out to cement its position on the emerging “diffusion LLM” frontier.
Sources
Back to AIPULSEN