Summary
Most AI labs are chasing intelligence benchmarks, but diffusion language models focus on speed. Sam Witteveen tests Inception’s Mercury 3 and Mercury Decide to see whether diffusion LLMs have finally caught up in quality. The video is sponsored by Inception, which provided early access.
He first explains how diffusion models differ from autoregressive ones. Instead of generating one token at a time like a typewriter, diffusion models start with a masked draft of the whole response and refine many tokens in parallel, which speeds up inference. Mercury 3 runs at over 1,000 tokens per second, and Inception’s benchmarks place it in the range of GPT-6 Luna and Claude Haiku 5.5, particularly on reasoning, instruction following and search. Inception has scaled parameters more than 40x from Mercury 1 Mini to Mercury 3, with intelligence scores up roughly 3x.
Mercury Decide is a fast decision model in the style of Jev, and it is currently free on OpenRouter. Independent evaluations from Vals AI show strong LegalBench results, around 87% versus 74% for Jev, at roughly a quarter of the cost, with median latency near 0.16 seconds per judgment. Sam also demonstrates a custom test app that replays token generation step by step and discusses using cheap, fast checks alongside a single heavy reasoning call.
📺 Source: Sam Witteveen · Published October 20, 2026
🏷️ Format: Benchmark Test
