Inception Labs releases Mercury 3 diffusion LLM at 1,068 tokens per second

By Splitfeed EditorialLast updated Oct 21, 2026, 15:00 UTC

On October 19, 2026, Inception Labs released Mercury 3, a diffusion language model the company says runs at 1,068 tokens per second. Inception claims that speed is four to eight times faster than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5. The company also says Mercury 3 outperforms both models on eight of twelve benchmarks, and it describes Mercury 3 as the most capable diffusion language model released to date. No independent benchmark results have been published to verify any of these figures.

The release matters beyond the speed numbers because diffusion language models represent a distinct architectural approach from the autoregressive transformers that dominate the current market. If the performance and speed claims hold up under independent testing, Mercury 3 would mark a meaningful competitive entry from a smaller lab against two of the most established names in the industry. The central dispute at this point is straightforward: Inception has published claims, but the figures come entirely from the company itself, which makes the benchmarks difficult to evaluate until external researchers or reviewers run their own tests.

None of the three camps, Pro-AI, Anti-AI, or Middle Ground, have published reactions to this story yet. Pro-AI observers would typically treat a speed improvement of this scale as evidence that competition is driving rapid gains and that diffusion architectures deserve serious attention alongside established approaches. Anti-AI voices would likely question whether self-reported benchmarks can be trusted and raise concerns about the broader pattern of labs making headline-grabbing performance claims without waiting for independent validation. The Middle Ground camp would probably call for caution, welcoming the architectural diversity while stressing that the numbers need third-party confirmation before they can be treated as settled.

The clearest thing to watch is whether independent researchers or reviewers publish their own evaluations of Mercury 3. If outside benchmark results arrive and align with Inception's claims, the competitive picture for diffusion models shifts substantially. If they diverge, the story becomes one about how performance claims are communicated at launch.