Kimi K3 Just Broke The Economics Of AI

Summary
Dr. Károly Zsolnai-Fehér of Two Minute Papers reviews Moonshot AI’s release of Kimi K3, an open-weights frontier multimodal model. He examines its coding and agentic capabilities—such as generating full 3D interactive web games, emulators, and UI clones—alongside its 2.8-trillion-parameter architecture and underlying technical innovations.

What is shown

• [00:00] Web and game coding demonstrations: Complex interactive software generated with Kimi K3, including a 3D browser-based RPG (Sworddream of the Nine Heavens), an interactive Game Boy Advance 3D emulator (GBArcade) running retro games [00:02], a functional macOS desktop clone with working apps [00:08], and an Animal Crossing–style 3D village game [00:13].
• [00:19] Benchmark evaluations: Coding comparisons across DeepSWE, Terminal-Bench 2.0, FrontierSWE, Program Bench, and SWE Marathon showing Kimi K3 competing closely with GPT-5.6 Sol, Claude Fable 5, Claude Opus 4.8, and GPT-5.5.
• [00:38] Concept-to-game agent workflow: A prompt taking a 2D sketch of a cowboy on a horse to generate 3D assets, Three.js terrain, and TypeScript logic.
• [00:46] Frontier model parameter timeline: A chart illustrating open frontier model sizes over time, placing Kimi K3 at the top at 2.8T parameters ahead of DeepSeek (1.6T), Xiaomi (1.02T), and Thinking Machines (975B).
• [01:01] Web interface: The Kimi chat interface showing options for Instant, K3, and K3 Swarm modes.
• [01:42] Architectural animations: Visual explanations of K3’s core architectural components:
  • Kimi Delta Attention (KDA) [01:42]: An attention memory mechanism metaphorically illustrated as a notebook selectively updated and faded over long context windows.
  • Attention Residuals [02:18]: Layer-to-layer version tracking allowing deep network layers to access earlier document representations.
• [02:58] Debugging and simulation projects: An automated debugging demo on a 3D horse-riding game, physics-based swinging mechanics [03:05], a 3D digital watch with 242 animated parts [03:13], and a simulated Gargantua black hole raymarcher [03:39].
• [03:29] Chip design & technical report: Highlights from the technical report (Kimi K3: Open Frontier Intelligence), including an automated chip design case study and local fluid simulation generation on Arch Linux [03:53].
• [04:26] Sponsored showcase: A demonstration of cloud GPU infrastructure through sponsor Lambda.

Claims & numbers

• Model size & architecture: The presenter notes Kimi K3 has 2.8 trillion total parameters (the displayed technical report confirms a mixture-of-experts model with 104 billion active parameters per token and a 1-million-token context window).
• Scaling efficiency: The presenter states the combination of Kimi Delta Attention and Attention Residuals delivers a roughly 2.5× improvement in scaling efficiency over Kimi K2 (delivering 2.5× more learning progress per compute budget rather than raw inference speed).
• Open weights: The presenter emphasizes that K3’s weights are released open and free to download permanently.
• Economics and API pricing: The presenter claims K3’s API is significantly cheaper than leading proprietary frontier models, exerting strong downward pressure on token prices across the industry.

Notable quotes

• “This is an open-weights model. Yes, we can download and own the weights for free, forever.” [00:32]
• “It is absolutely, stupendously, humongously big: 2.8 trillion parameters.” [00:52]
• “Hold on to your papers, fellow scholars, because it results in a 2.5x improvement in scaling efficiency over Kimi K2.” [02:58]

Assessment
This is a third-party educational overview and review video summarizing Moonshot AI’s technical paper and official showcase demos. The demonstrations feature highlight reels and curated community and developer examples rather than unedited, live-prompted generation sessions on camera.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.