Mercury 3 shows diffusion LLMs scale: 40x bigger, 3x smarter, still 1000 tok/s
volokuleshov · x · 2026-10-20
Inception released Mercury 3, which it calls the biggest and smartest diffusion LLM to date, still running at 1000 tokens/sec.
- Key claim: diffusion scales. With 40x the total parameters of Mercury 1 (mini), intelligence across standard benchmarks rose roughly 3x, suggesting the scaling laws behind autoregressive progress hold for diffusion models too.
- The company's chart traces five Mercury generations: each beat contemporaries like Gemini Flash Lite, Claude Haiku 4.5 and GPT-5.6 Luna on speed by 4-10x, with M3 claiming 8x faster than GPT-6 Luna.
- The intelligence score averages GDPval-AA v2, SciCode, GPQA Diamond, Terminal-Bench v2.1, BrowseComp+ and more.
If the numbers hold, it's a strong rebuttal to the view that diffusion LM is a small-model-only paradigm.
More from Models
- Anthropic prepping new model that infers intent from minimal input, alongside Gemini 4 Carbon and OpenAI Terra — dotey · 2026-10-20
- User slams Xiaomi MiMo over hard-to-cancel $60 subscription — bookwormengr · 2026-10-20
- Smaug Open-Sources SOTA AI Image Detector, Hits 86.7% Recall at 5% False Alarms — bindureddy · 2026-10-20
- Anthropic Email Reveals Claude Mythos 5.1, Opus 5.5, Sonnet 5.5 in Expanded Cyber Verification Program — Sauers_ · 2026-10-20
- Dream-1 launches: 41% of subtasks in 5 minutes, outpacing top LLM agents — gharik · 2026-10-20
- Xiaomi's MiMo-V2.6-Pro is 'the first LLM that thinks like a teenager' — and it's actually good — hrishioa · 2026-10-20