Inception Releases Ultra-Fast LLM 'Mercury 2.5' Generating 770 Tokens Per Second

Inception has released its proprietary LLM Mercury 2.5. According to Artificial Analysis, this model generates 770.4 output tokens per second, a significantly faster speed compared to the median (108.6 t/s) of reasoning models in the same price range.

Performance and Cost

• Speed: Generating 770.4 tokens per second ranks it 2nd among 175 models, but its Time to First Token (TTFT) is 2.91 seconds, slightly slower than the median (2.22 seconds).
• Intelligence: The Artificial Analysis Intelligence Index v4.3.2 score is 12 points, lower than the median for the same price range (13 points), placing it 91st among 175 models.
• Cost: Priced at $0.25 / 1M tokens for input and $0.75 / 1M tokens for output. The blended price with a 90% cache discount is $0.14 / 1M tokens.
• Conciseness: During the Intelligence Index evaluation, it generated a total of 35M output tokens, showing very concise responses compared to the median for the same price range (85M).

Technical Specifications and Deployment

• Architecture: It is a reasoning model (supporting extended thinking) with a 260k token context window. It supports only text input/output and does not support multimodal (image input).
• Deployment: Currently accessible only via the Inception API, with weights and parameters remaining private.