Learn about this first
Inception Labs releases diffusion-based large language model Mercury 3, saying it reaches 1,068 tokens/sec on NVIDIA GPUs, has a 1M-token context window, and has over 40 times as many parameters as Mercury 1. Inception says the model leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 benchmarks and is about 4 to 8 times faster than the two models. It is now available on OpenRouter via API only; weights and quantized versions are not available.
Pricing is $0.20 per million input tokens and $0.70 per million output tokens. For the launch period, prices are 50% off: $0.10 per million input tokens and $0.35 per million output tokens. Cached tokens receive an additional 90% discount off the input price, and new accounts get 100 million free tokens. Inception has not said when the discount ends.
Artificial Analysis clarified that the AA-Omniscience and GDPval-AA v2 scores cited in Inception's release graphic are not its Intelligence Index scores, do not follow its methodology, and cannot be directly compared with scores on Artificial Analysis: AA-Omniscience uses a private dataset, with only a 10% public subset open-sourced, while GDPval-AA v2 is comparable only when its scoring method and model comparison set are used.
AI-generated from reports · updated 26 days agoLatest developments
Inception Releases Mercury 3 Diffusion Language Model, Claiming to Surpass GPT-6 Luna and Claude Haiku 5.5 on Some Benchmarks
Event progress3
Clarification: Scores such as AA-Omniscience are not comparable
Artificial Analysis: Artificial Analysis clarifies that the evaluation scores in the Mercury 3 launch cannot be directly compared with its Intelligence IndexMercury 3 model launches on OpenRouter
OpenRouter: Inception's diffusion model Mercury 3 launches on OpenRouterInception Labs Releases the Mercury 3 Diffusion Model
Inception Labs: Blog: Inception Labs releases diffusion-based large language model Mercury 3, reaching 1068 tokens/sec