Build Anything with Kimi K3, Here’s How

Summary
In this video, presenter David Ondrej provides an overview, technical breakdown, and hands-on tutorial for Kimi K3, a 2.8-trillion-parameter open-weights multimodal model released by Chinese AI company Moonshot AI. He explores K3’s benchmark performance across coding, 3D design, and legal tasks against frontier models like Claude Fable 5 and GPT-5.6 Sol, analyzes its novel attention architecture, and demonstrates its capabilities live inside the Pi and Kimi Code agent harnesses.

What is shown

• Benchmark comparisons and architecture slides: Analysis of Frontend Code Arena, SWE Marathon, Harvey LAB-AA legal evals, parameter size comparisons (2.8T), and the Kimi Delta Attention / Attention Residuals mechanism [00:04 - 03:15, 08:48 - 10:25].
• Hardware self-hosting breakdown: Cost and component estimate ($16k–$22k) for running 2-bit/3-bit quantized Kimi K3 locally at 5–10 tok/s on refurbished dual-EPYC enterprise servers [10:26 - 14:20].
• Installing Kimi Code CLI: Downloading and installing Kimi Code via curl command in the terminal [16:42 - 17:04].
• Prompting Kimi K3 in the Pi harness: Kicking off two parallel coding tasks: rebuilding Blender as an in-browser 3D web app [18:10] and redesigning his personal hiring landing page with live follower counts [18:43].
• Interactive Blender clone demo: Opening and testing the generated 3D modeling web application in the Brave browser, verifying mesh scaling, rotation, translation, and edit mode [20:16 - 21:44].
• Landing page inspection: Reviewing the generated Next.js landing page with live-updated subscriber counts and UI animations [22:45 - 23:30].
• Kimi Code CLI legal analysis: Pointing Kimi Code with maximum thinking effort at a 41-page legal PDF (Apple v. OpenAI), with Kimi evaluating claims, defense counterarguments, and predicting settlement/trial odds [26:17 - 29:35].

Claims & numbers

• Model specifications & release: The presenter states Kimi K3 is developed by Moonshot AI, has 2.8 trillion total parameters (making it the largest open-weights model to date), and its weights arrive on July 27, 2026 under a Modified MIT license [01:43 - 02:01, 04:46].
• Benchmarks cited:
  • SWE Marathon: Kimi K3 scored 42.0%, beating Claude Opus 4.8 (40.0%), GPT-5.6 Sol (39.0%), and Claude Fable 5 (35.0%) [00:04].
  • Frontend Code Arena: Kimi K3 leads with an Arena score of 1,679, ahead of Claude Fable 5 (1,674) and Claude Opus 4.7 (1,640) [00:21].
  • FrontierSWE: Kimi K3 scored 81.2%, behind Claude Fable 5 (86.6%) but ahead of GPT-5.6 Sol (71.3%) [01:00].
  • Terminal Bench 2.1: Kimi K3 scored 88.3%, just behind GPT-5.6 Sol (88.8%) and ahead of Claude Opus 4.8 (84.6%) [01:03].
  • Harvey LAB-AA (Legal all-pass rate): Kimi K3 achieved 26.7%, nearly double Claude Fable 5 at 14.2% [03:07, 26:49].
  • DesignArena 3D Design: Kimi K3 ranks #1 with a 1288 Elo score [06:47].
• Pricing: API pricing for Kimi K3 is stated at $3.00 input and $15.00 output per 1M tokens, which the presenter notes is three times cheaper than Claude Fable 5 ($10 / $50 per 1M tokens) [04:20 - 04:30].
• Architecture: The presenter explains K3 replaces traditional residual stream scratchpads with “Attention Residuals” and uses “Kimi Delta Attention” (KDA) linear attention (three KDA layers followed by one full-attention layer), yielding up to 6.3x faster decoding for 1M-token contexts [08:48 - 10:24].
• Chip design claim: The presenter claims K3 autonomously designed and verified a chip architecture for a nano model in a 48-hour run [15:00 - 15:25].
• Task demo cost: Building the full web-based Blender clone cost $1.30 in API tokens [19:42].

Notable quotes

• [00:45] “And this is the first time ever where we have an open-source model crushing the best closed-source models out there.”
• [05:40] “What matters is the cost per task. If we look at DeepSWE, we can see that Sonnet 5 is the single most expensive model right here. Why? Because to complete the same task... it requires way more runs, way more turns.”
• [23:33] “Rather than remembering syntax, you need to be a better architect and know how to guide these models.”

Assessment
This video is a sponsored hands-on review and tutorial by David Ondrej covering Moonshot AI’s Kimi K3. The live demonstrations in terminal agents (Pi and Kimi Code) to build functional browser apps and review documents are genuine, though the external benchmarks and 3D web examples are presented from third-party social posts and promotional test runs.

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