Inception Launches Mercury 3 Diffusion Model with Million-Token Context and High Speed

Inception Labs has launched Mercury 3, a next-generation diffusion language model marking a key breakthrough in scaling non-autoregressive text generation. The model achieves over 1,000 tokens per second while expanding the context window to 1 million tokens. Notably, it scales up by 40 times compared to its predecessor without sacrificing speed, validating the scalability of diffusion architectures for large-scale, real-time applications.

Cross-referenced reports indicate that Mercury 3 is now live on OpenRouter with limited-time discounts. Official claims state it outperforms GPT-6 Luna and Claude Haiku 5.5 on benchmarks like IFBench, GPQA Diamond, and SWE-Bench Atlas. However, Artificial Analysis clarifies that cited scores rely on private datasets or specific derived metrics, lacking direct comparability with standard Intelligence Index rankings. The Decoder notes that while inference speed improves significantly, overall performance varies, with distinct advantages in instruction-following tasks.

This release impacts the developer ecosystem by proving diffusion models as efficient alternatives to traditional autoregressive LLMs, particularly for low-latency, high-throughput scenarios. Despite debates on benchmark comparability, Mercury 3’s technical path offers new optimization directions, potentially accelerating the adoption of parallel generation in code assistance and long-document processing.