Inception launches Mercury 3 with parallel text generation

Inception Labs has introduced Mercury 3, a diffusion language model that generates tokens in parallel and refines responses through successive passes rather than producing text sequentially. The company reports throughput of 1,068 tokens per second on NVIDIA GPUs, positioning the model for browser agents, customer support, live search and interactive education. The GPU configuration was not disclosed, and application performance will also depend on network latency, tool calls and workload characteristics.

Inception says the model leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 evaluations at medium settings. Its reported strengths include instruction following, web research and software engineering, although it trails competitors on some coding and context-retrieval tests. These results are vendor-reported rather than independently confirmed. Access is API-only, with no open weights or disclosed parameter count, limiting options for teams requiring self-hosting.

The OpenAI-compatible API supports tool calling, structured outputs and adjustable reasoning effort. Access is available through Inception’s platform, OpenRouter, Baseten and Vercel. Launch pricing includes $0.10 per 1M input tokens and $0.35 per 1M output tokens, with 100M tokens available free to new accounts. The 50% discount has no published end date, making standard pricing the safer basis for cost planning. Teams considering adoption should validate latency and accuracy on their own workloads.