Inception releases Mercury 3, a diffusion LLM claiming 1,068 tokens per second and a 1M-token context
- Sources: primary, discussion
- Summary: Inception says Mercury 3 is more than 40x larger than Mercury 1, generates 1,068 tokens per second on NVIDIA GPUs, and has a 1M-token context, up from 260K. It is OpenAI API compatible and 50% off at launch. Inception claims parity with GPT-6 Luna and Claude Haiku 5.5 but the post gives no evaluation method.
- Why it matters: Latency-bound agent loops get a non-autoregressive model to test, though the quality claims are vendor-only for now.