Inception released Mercury 3, the newest version of its diffusion language model, which generates text in parallel instead of one token after another.
The company reports 1,068 tokens per second on Nvidia GPUs and a context window of 1 million tokens.
That throughput figure is the pitch. Autoregressive models decode serially by construction, so the only ways to go faster are batching requests or drafting tokens with a smaller model and verifying afterwards. Diffusion sidesteps the ordering entirely by denoising a whole block at once.
Speed is not the same as quality
Earlier diffusion language models traded accuracy for throughput, and the approach struggled on tasks that need strict left-to-right dependencies such as long arithmetic or code with matching braces.
Inception’s own benchmark table shows Mercury 3 leading on the metrics the company chose to publish. Those numbers are self-reported and were not run by an independent evaluator.
The commercial target is clear. Agentic workloads make many short model calls in sequence, and latency compounds across every hop. A model that returns a full block in one pass changes what an agent loop can afford to attempt.
Whether diffusion decoding holds quality at a million tokens of context is the open question. Long-context work depends on precise recall of earlier material, and denoising a distant block from a noisy start is a different problem than continuing a sequence.
