Inception Labs has released Mercury 3, a diffusion language model the company reports at more than 1,000 tokens per second with a 1 million token context window.
The architecture is the point. Almost every language model in production is autoregressive - it writes one token, looks at what it wrote, then writes the next, which means output speed is bounded by a sequential loop. A diffusion model starts from noise and refines an entire draft at once, in parallel, the way image diffusion models do. That removes the sequential bottleneck, which is where a four-figure token rate comes from rather than from faster hardware.
The honest caveat: diffusion LLMs have historically traded quality for speed, and Inception has not published comparative benchmark results against frontier autoregressive models alongside the speed figure. Treat 1,000 tokens per second as a throughput claim rather than a capability one until there are third-party evaluations. If it holds up at usable quality, the applications are the ones where latency is the product - live voice, agent loops that make many calls, and anything where a user is watching a cursor blink.