Forty times larger than the first Mercury but just as fast, Inception’s new diffusion model claims to outperform GPT-6 Luna and Claude Haiku 5.5.
The Inception start-up presented on 19 October Mercury 3, the new generation of its broadcast language model. Unlike self-regressive models, which write one token after another, this architecture produces and refines several tokens in parallel, as image generators. The model has more than 40 times more parameters than Mercury 1, with a level three times higher on the usual test benches, while maintaining the same flow rate: 1,068 chips per second on GPU Nvidia.
Fast, and cheaper than small models
Inception states that Mercury 3 exceeds GPT-6 Luna and Claude Haiku 5.5 on a series of benchmarks, turning four to eight times faster. The context window is up to a million chips, compared to 260,000 for Mercury 2.5. The launch price, reduced by half, is $0.10 per million tokens in entry and $0.35 in exit (0.20 and $0.70 at the normal rate).
Compatible with the OpenAI API, the model is available on the Inception platform as well as via OpenRouter, Baseten and Vercel. The company is dedicated to uses where latency is added: navigation and research agents, voice assistants, customer support or programming sub-agents.
Read the full article on the original source
Go to the source →