Inception Labs ships Mercury 3: diffusion LLM beats GPT-6 Luna at 1,068 tokens/sec
volokuleshov · x · 2026-10-20
Inception Labs released Mercury 3, the most capable diffusion-based LLM to date, showing diffusion scaling laws hold:
- Scale: 40x larger than Mercury 1, 3x smarter on standard benchmarks — autoregressive-style scaling works for diffusion.
- Quality: on par or better than GPT-6 Luna and Claude Haiku 5.5 (GPQA Diamond 89.23, SWE-Bench Atlas 51.08, BrowseComp+ 55.18).
- Speed: 1,068 tokens/sec on NVIDIA GPUs, 4–8x faster than peers; 1M-token context (up from 260K).
- Price: $0.20/$0.70 per 1M input/output tokens, 50% off at launch; cached tokens 90% off.
- Early partners use it for voice support, tutoring, browser agents, and data engineering.
More from Infra
- AMD: World now consumes 40 quadrillion inference tokens monthly, up 95x in two years — Beth_Kindig · 2026-10-20
- Meta details Prometheus gigawatt-scale supercluster, targeting ~1M GPUs — BenBajarin · 2026-10-20
- Microsoft calls Maia 300 its biggest silicon bet, multi-gigawatt deployment from 2027 — pdamodaran · 2026-10-20
- Naive torch.matmul uses ~1% of your GPU: a series on kernel engineering — Winter_Mistake_3185 · 2026-10-20
- NeurIPS 2026 paper 'Intelligence per Watt' measures on-device AI efficiency across 1M queries — HazyResearch · 2026-10-20
- Inception Labs launches Mercury 3, a diffusion LLM claiming 1000 tok/s inference — cmrdporcupine · 2026-10-20