Inception Launches Mercury 3 Diffusion Model with Million-Token Context

Inception Labs has officially launched Mercury 3, a next-generation diffusion language model that marks a significant breakthrough in scaling text diffusion architectures. The model expands the context window to 1 million tokens while maintaining generation speeds exceeding 1,000 tokens per second, even as its scale increases 40-fold compared to previous versions, validating the feasibility of non-autoregressive models at large scales.

Cross-verified reports indicate that Mercury 3 performs exceptionally well on benchmarks such as IFBench, GPQA Diamond, and SWE-Bench Atlas, with official claims stating it outperforms GPT-6 Luna and Claude Haiku 5.5. This achievement breaks the traditional bottleneck where diffusion models struggle to balance speed and quality, proving their competitiveness in complex reasoning tasks. Currently available via OpenRouter with a limited-time 50% discount, the model lowers the barrier for developer adoption.

The release of Mercury 3 profoundly impacts the AI infrastructure landscape. It demonstrates that diffusion models can serve as efficient alternatives to autoregressive models, particularly in scenarios requiring real-time responses and long-context processing. This provides developers with new technical options and accelerates the convergence of multimodal and pure text generation technologies, driving the industry toward higher throughput and lower latency.