Inception: every model, side by side

Inception's diffusion language model generates and refines tokens in parallel instead of one at a time. It reaches a reported 1,107 tokens per second on standard GPUs, with a 260K-token context window and a claimed 10-point intelligence gain over Mercury 2.

Inception's Mercury 2, the first reasoning diffusion language model. Instead of emitting tokens one at a time it produces and refines many in parallel, clearing 1000 tokens per second on standard GPUs, over 5x faster than Claude 4.5 Haiku and GPT-5 Mini. 128K context, native tool use, schema-aligned JSON.

The first diffusion large language model. Discrete diffusion runs 5 to 10x faster than speed-tuned models like GPT-4.1 Nano and Claude 3.5 Haiku while matching them, which is what makes voice agents and search viable.