Kimi K3 Has 2.8 Trillion Parameters and Uses 1.8% of Them Per Token. Here Is the Mathematics of Why That Is Genius.
Last Updated on July 23, 2026 by Editorial Team
Author(s): Dr Swarneendu AI
Originally published on Towards AI.
Kimi K3 Has 2.8 Trillion Parameters and Uses 1.8% of Them Per Token. Here Is the Mathematics of Why That Is Genius.
Two weeks ago the largest open-weight model in existence had one trillion parameters.
After explaining why sheer model size isn’t deployable, the article walks through how Kimi K3 makes a 2.8T model practical using four interlocking engineering decisions: a sparse Stable LatentMoE (896 experts with only 16 active per token) plus Quantile Balancing to avoid MoE routing collapse without a tunable hyperparameter; Kimi Delta Attention to enable a 1M-token context by replacing quadratic attention with a mostly-linear hybrid mechanism; Attention Residuals to overcome depth-related gradient and information propagation bottlenecks via multi-layer skip retrieval; and MuonClip (Muon optimizer + QK-Clip, including per-head Muon) to prevent training catastrophes like loss spikes and attention-logit overflow. It then contrasts K3’s real-world benchmark strengths—especially for agentic, long-context, parallel workloads—with where frontier single-agent reasoning can still lead elsewhere, and closes with the “cost arithmetic” showing K3 can be dramatically cheaper for practical token-heavy workflows while staying close enough in performance to matter.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
