2.8 Trillion Parameters - Kimi K3 is here

2026-07-16 · Source: 1littlecoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

Moonshot has released Kim K3, a new open-source large language model featuring an unprecedented 2.8 trillion parameters, making it the largest open-source model to date. This MOE model incorporates innovations like Kimmy Delta attention (KDA) and attention residuals, offering a vast 1 million context window and multimodal capabilities for vision and video input. Available live as Kimmy K3 Max, it demonstrated superior front-end website generation compared to OpenAI's output in a specific test, albeit taking 15 minutes versus OpenAI's 4 minutes 28 seconds. Moonshot claims Kim K3 delivers "frontier level performance," ranking second only to Claude Fable 5 and GPD 5.6 6 soul. Its pricing is approximately 3-5 times higher than Kimik K2.7, with input at \$3 per million tokens and output at \$15 per million tokens, aligning with OpenAI's Saul 5.6.

Key takeaway

For AI Scientists and Machine Learning Engineers evaluating large-scale open-source models, Kim K3 presents a significant new option. Its 2.8 trillion parameters and 1 million context window offer unprecedented scale for complex tasks, particularly in multimodal and web generation applications. You should consider testing Kim K3 Max for projects requiring advanced front-end development or extensive context. Be aware of its higher cost compared to previous versions and potentially slower inference times due to its diligent processing.

Key insights

Kim K3 sets a new open-source scale benchmark with 2.8 trillion parameters and advanced web generation capabilities.

Principles

• Large-scale MOE models can achieve frontier performance.
• Proprietary attention mechanisms enhance model capabilities.
• Open-source models can compete at the trillion-parameter scale.

Method

Kim K3 utilizes internal computer use skills, bash, and a self-correction loop (screenshot, QA, fix) to generate complex, aesthetically brilliant websites, taking 15 minutes for a detailed output.

In practice

• Deploy Kimmy K3 Max for advanced front-end generation.
• Explore Kimmy K3 Swarm Max for agent deployments.
• Integrate multimodal inputs for diverse applications.

Topics

• Kim K3
• Large Language Models
• Open-Source AI
• Multimodal AI
• Website Generation
• Moonshot AI
• Mixture of Experts

Best for: AI Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• BREAKING: China's Kimi K3 Just Dropped a 2.8 Trillion Parameter Open AI Model — Kimi K3 introduces a 2.8T-parameter open MoE model with a 1M-token context, setting a new scale benchmark.
• I tested Kimi K3 with INSANE prompts.... — Kimi K3, a 2.8T parameter model, excels at complex multimodal tasks, demonstrating strong instruction following and "out-of-distribution" thinking.
• Kimi K3 Is Now Open-Weight: The Full Breakdown of the 2.8T Frontier Model — Moonshot AI's Kimi K3 is the largest open-weight, native multimodal model, setting a new benchmark for frontier AI.
• Kimi K3 just beat FABLE. — Chinese open-source models, like Kimi K3, are now frontier-class competitors to proprietary LLMs.
• Kimi K3 Is the Biggest Open Source Model Ever. Almost No One Can Run It. — Kimi K3's massive 2.8 trillion parameters make self-hosting impractical despite its open-source nature, highlighting a compute barrier.
• Kimi K3 Explained: 2.8 Trillion Parameters, 16 Active Experts, 1 Huge AI Shift — Kimi K3 redefines AI scale by activating only a small fraction of its 2.8 trillion parameters per token, shifting focus to routing efficiency.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by 1littlecoder.