Inception unveils Mercury 3 diffusion‑based LLM, claiming Luna‑level quality at 4‑8× speed (~1,000 tokens per second)
claude
| Source: Techmeme | Original article
Inception unveiled Mercury 3, a diffusion‑based large language model that it says matches Luna and Haiku performance while delivering 4‑8× speed, producing about 1,000 tokens per second.
Inception, the California‑based AI research and product firm, unveiled Mercury 3 on 19 October 2026, branding it the most capable diffusion‑based large language model to date. The company says Mercury 3 matches the quality of leading models such as GPT‑6 Luna and Claude Haiku 5.5 while delivering inference speeds four to eight times faster – roughly 1,000 tokens per second on NVIDIA GPUs – and supporting a 1 million‑token context window. Launch pricing was set at $0.10 per million input tokens and $0.35 per million output tokens.
The announcement marks a shift in the industry’s scaling strategy. Most competitors rely on dense transformer architectures that demand ever‑larger data centres to meet latency expectations. By leveraging diffusion techniques, Inception claims to cut compute cycles dramatically, potentially lowering the energy and hardware costs of running high‑throughput language services. If the speed advantage holds in independent tests, developers could build more responsive agents, real‑time translation tools, and interactive assistants without the expense of massive GPU farms.
Stakeholders will be watching three key developments. First, third‑party benchmarks will verify the 1,068 tokens‑per‑second figure and the claimed parity with Luna and Haiku on standard evaluation suites. Second, adoption rates will reveal whether the modest per‑token pricing and speed gains translate into tangible cost savings for enterprises and cloud providers. Finally, rival labs are likely to respond with their own efficiency‑focused releases, prompting a new wave of competition centered on inference speed rather than sheer model size. Mercury 3 could therefore become a bellwether for the next generation of cost‑effective, high‑performance LLMs.
Sources
Back to AIPULSEN