I Tried Kimi K3 for a Week. Here’s What Happened.

A developer tested Kimi K3, a new open AI model from Moonshot AI, for a week of real work. The model handled a 1M token context window by refactoring a 40,000-line codebase, and ranked #1 on Arena's Frontend Coding leaderboard and nextjs.org/evals. While not the best model overall, it is the best open model for frontend work at ~35% the cost of Claude Fable 5, with full open weights expected by July 27.

A few days ago I posted my first impressions of Kimi K3 on LinkedIn https://www.linkedin.com/pulse/kimi-k3-strong-new-chinese-ai-model-milad-joodi-7vxbe/?trackingId=it5Ujm9OT2CmxjSCu9CHJw%3D%3D . Since then, I’ve been using it daily for real work — not benchmarks, not demos. Here’s what actually held up. The Context Window Isn’t Marketing I tested the 1M token claim the hard way: dumped a 40,000-line codebase into it, then asked it to refactor a utility function that was referenced across 12 files. It traced every call site correctly. No “remind me what file X does” loops. That’s the difference between reading specs and actually using it. Frontend Coding: The Hype Checks Out Kimi K3 is 1 on Arena’s Frontend Coding leaderboard right now, ahead of Claude Fable 5. I was skeptical, so I gave it a real task: build a responsive dashboard with real-time charts, dark mode, and collapsible sidebar. The code was clean, state management made sense, and it handled edge cases I hadn’t explicitly asked for. Not perfect, but genuinely competitive. It’s also 1 on nextjs.org/evals — the first open model to beat all proprietary ones on that benchmark. Agentic Work: Good, Not Magic I ran a multi-step workflow: read design tokens, generate components, write tests. It handled the chain without losing context. When I changed a token midway, it adapted. But let’s be real — it’s not autonomous. I still had to steer. It’s more like a really good pair programmer than a self-driving coder. On Agent Arena, it’s currently 4, matching Claude Opus 4.8 and GPT-5.6 Sol. Strong, but not 1. The Honest Truth Moonshot themselves say K3 “still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol.” That’s fair. It’s not the best model in the world. But it’s the best open model for frontend work right now, and at ~35% the cost of Fable 5 for similar software engineering tasks, the value is hard to ignore. Why the Open Weights Matter Full weights dropping soon targeting July 27 . For a 2.8T parameter model with this context window, that’s not just “open source” — that’s a foundation people will actually build on. Fine-tune it, run it locally, integrate it into your stack. That’s the part that changes the game. Bottom Line Kimi K3 isn’t perfect. It won’t replace Claude or GPT for everyone. But it’s the first open model where I didn’t feel like I was trading capability for accessibility. It just works. If you want to try it: kimi.com P.S. — Original first impressions post .