Mercury 3 argues text diffusion scales like autoregression

Most large language models write one token at a time, left to right. Inception builds diffusion models instead, which refine a whole block of text in parallel, and its new Mercury 3 is the clearest case yet that the approach scales. Inception says Mercury 3 has more than 40 times the parameters of its first Mercury model and roughly 3 times the benchmark score, which is the real claim here: the scaling laws that made autoregressive models better with size appear to hold for diffusion as well. That was the open question after Mercury 1 proved such models could be built commercially.

The practical draw is speed. Inception reports 1,068 tokens per second on Nvidia GPUs, 4 to 8 times faster than speed-optimized frontier models like GPT-6 Luna and Claude Haiku 5.5, while claiming quality on par or better. On a browser-agent test using 50 live Mind2Web navigation tasks, Mercury 3 completed 50%, which Inception says beat GPT-6.1-Sol at similar token volume and lower per-token cost. The model has a 1 million token context window, up from 260K in Mercury 2.5, and is priced at $0.20 per million input tokens and $0.70 per million output. It is OpenAI API compatible, so it can drop into an existing stack. Nearly all the comparison numbers come from Inception or early partners, and the post gives little independent benchmark detail, so treat the quality claims as unverified.

Why it matters

If you run latency-sensitive work like voice assistants, live search, or agent loops that make many model calls, a model this fast at competitive quality could cut both response time and cost. The speed is the selling point, so the thing to test on your own workload is whether quality actually holds at that pace, not the headline tokens-per-second.