Two-Day SaaS with Kimi K3: Is Claude Really Dethroned?

The narrator introduces himself as an unemployed ex-big-tech engineer with 25 years in the industry who built an automation that cuts short clips from his long vlogs and pushes them to social platforms; after months of silence, a few clips take off on TikTok, Facebook and Instagram and bring a real wave of followers. The detail that hooked me is the one-link rule on profiles: on Instagram, anything below the fold might as well not exist. The opening frame states the bet plainly, with the host in a home office in front of a code monitor next to a big GOODBYE CLAUDE question.

When the new audience wants one page with everything, he tries ready-made bio-link services such as Linktree, Beacons and Squarespace; the supposedly free pages stamp their own branding everywhere, then bury the content under upgrade emails and custom-domain pressure. So he flips the rule and decides to write his own link-page generator, picking Moonshot's open-weight Kimi K3 as the engine. The model is said to match or beat leading American closed models on many measures, and it enters as the no-single-vendor alternative to the Claude Fable he uses for daily work.

With no giant home cluster, he reaches Kimi K3 through OpenRouter, paying a reseller that runs the model in data centers inside the United States rather than paying Moonshot directly. This seemingly small detail carries one of the video's theses: open weights mean choosing who gets paid and where the work runs. The published model cards describe Kimi K3 as a sparse mixture-of-experts design with 2.8 trillion total parameters and a 1M-token context window.

He then explains the agent harness in plain terms: the context window is the model's short memory, and once full, older details fall away and fabrication starts; the harness manages that memory with global and project-level memory files, compacting and restarting from the same baseline when needed. The rig also gives the model tools and skills; reading data, opening a browser or calling a remote service are tools, while repeatable task recipes in Markdown files are skills. Most tools follow the Model Context Protocol so they can be reused across rigs, and he warns that downloaded skills can carry hostile lines, which is why he writes his own.

Every major lab has a rig tuned for its own model: Claude Code on the Anthropic side, Kimiko on the Moonshot side, Codex on the OpenAI side, plus open-source rigs such as OpenCode, PiCode and Hermes that work with many models. He deliberately picks what he calls the most boring option, the Visual Studio Code Agent rig, because he wants a fair track: he tested Claude 5 Fable on the same rig when it launched and believes it gives no model a hidden edge. That neutrality claim underpins the whole comparison.

For the first job he asks Kimi for a technical design document; his prompt states the end-to-end requirements, demands an auto-scalable stack a solo developer can maintain, and explicitly asks the model to ask questions. In that design talk, Kimi asks more questions than Claude 5 Fable did, digging into functional and technical details, which he reads as a good sign. The skeleton of the architecture emerges after about twenty minutes.

He pushes back on two points in the proposal: keeping every bio page's view data in the same database as the main app could slow the whole system on a traffic spike, and self-run PostgreSQL is harder to maintain than a managed database service. An old build-versus-buy debate with former coworkers reopens, and he jokes that talking with AI feels like shadows cast on a cave wall. The agreed stack ends up as React plus Tailwind on the front, Next.js in the middle, Convex for data, Google Analytics for measurement, Clerk for login and Stripe for payments.

For coding he skips the common pattern of slicing work across subagents; since smaller models are usually weaker, he wants to see Kimi write the code itself. The order is crisp: advance phase by phase from the design paper, write unit tests each time plus end-to-end tests for critical flows, run them, fix what breaks, loop until green. He grabs a coffee while it works; the build runs over an hour and the model twice falls into a strange deadlock, burning meter without progress until the rig steps in after five minutes. He reads that as a model-rig integration flaw.

At validation he does bleak math: a human can review only about 500 lines a day at high quality, yet the model produced thousands; a weeks-long job gets triaged by a code-review skill that maps defects by severity so scarce human attention lands on the worst areas. After hours of review he runs the app locally; apart from minor style issues the core behavior is surprisingly accurate. Not flawless in one shot, but comfortably above a pass.

The last snag is a hook-ordering race; Kimi struggles with it and suddenly asks for the production deploy keys to try a fix live. His joke is sharp: a top-tier model asking for the house keys on the first date. He refuses, tries a few more prompts until the defect is fixed, and the Vercel deploy lands cleanly; idea to production takes under two days.