Inception has introduced Mercury 3, a diffusion-based language model that the company claims outperforms GPT-6 Luna and Claude Haiku 5.5 on several benchmarks. The model generates 1,068 tokens per second on Nvidia GPUs, significantly faster than the 240 tokens per second for Claude Haiku 5.5 and 130 tokens per second for GPT-6 Luna.
The architecture allows the model to generate multiple words simultaneously, unlike standard autoregressive models. Mercury 3 achieved 76.84 points on IFBench and 51.08 points on SWE-Bench Atlas. While it leads in instruction following and autonomous coding, Claude Haiku 5.5 performs better in long-text information retrieval and specific coding tasks.
Inception reports that Mercury 3 has over 40 times more parameters than its predecessor, Mercury 1, while maintaining the same speed. The model supports a context window of up to 1 million tokens. It is currently available via Inception's platform, OpenRouter, Baseten, and Vercel, utilizing an API compatible with OpenAI standards.
Pricing is set at $0.20 per million input tokens and $0.70 per million output tokens, with a temporary 50% discount at launch. Early adopters, including Octen AI and General Learning, report significant latency improvements in search and educational applications. The startup previously raised $50 million from investors like NVentures, M12, and Snowflake Ventures.
