A 2.8 trillion parameter model that anyone can download just beat the best closed models at front-end coding. Not by a hair. By thirteen percentage points. This breakdown comes from Matthew Berman, who covered the Kimi K3 release from Moonshot AI while literally sweating through 100 degree heat to record it, and I think it’s one of the clearest takes I’ve seen on what this release actually means.
I’ll be honest, my first reaction was skepticism. We’ve all seen the benchmark hype cycle. But the creator digs past the headline numbers into pricing, token efficiency, and the caveats that most coverage skips entirely. That’s what makes it worth your time.
🚀 The core claim
Moonshot AI, a Chinese lab, released Kimi K3 as an open weights model. On Arena AI’s front-end development benchmark, it scored 76% versus 63% for the number two model. That puts an openly available model above the closed frontier labs on a specific, meaningful task for the first time.
The specs the author walks through:
- 2.8 trillion parameters, the biggest open model released so far
- 1 million token context window
- Built for long horizon coding, knowledge work, and reasoning
- Released fully open, including the algorithmic details behind how they built it
And no, you can’t run this on your laptop. This is a datacenter model. But you can rent it, self-host it if you have the hardware, and inspect exactly how it was made.
The demo video Moonshot published is the part that stuck with me. The creator points out that Kimi K3 edited that video itself. In the demo it builds out a 3D simulated world with real-time reflections and a dynamic daylight cycle. It looks like a real game engine scene, not a toy.
💰 The pricing story is more complicated than it looks
This is the section I’d have missed on my own, and it’s the sharpest part of the analysis.
On paper, Kimi K3 runs about $3 per million input tokens and $15 per million output. That’s roughly half the price of the leading closed model. Sounds like an easy win.
But the expert pulls up the DeepSWE benchmark, which plots completion success rate against average cost per task. And there, Kimi K3 sits at essentially the same cost per task as the closed competitor it undercuts on sticker price. Around $4.70 either way.
Why? Intelligence density. Kimi K3 burns roughly twice the tokens to finish the same job. Half the price per token, double the tokens, same bill. The lesson here applies way beyond this one model: never evaluate an LLM on price per token alone. Price per completed task is the number that hits your card.
The author also notes Kimi K3 is genuinely slow. He kicked off a Rubik’s Cube simulator build as a live test and it ran past 30 minutes before finishing. It did finish, and the result worked cleanly with proper scrambling and solving, but token hungry and slow is a real tradeoff for anything latency sensitive.
📊 The independent signals that back it up
What convinced me this isn’t just benchmark gaming is that people outside Moonshot are seeing the same thing on their own evals.
- Vercel’s CEO reported Kimi K3 as the top performer on their Next.js evaluations, hitting a 92% success rate with agents.md, ahead of every proprietary model on that comprehensive web engineering benchmark. First time an open model has led that board.
- An editorial writing benchmark moved Kimi K3 from 21st place to first for writing in their house voice, at 2840 ELO. The team behind it noted it’s five times cheaper than the model it displaced.
- David Sacks, the US AI czar, publicly flagged the result as concerning, which tells you the release registered at the policy level, not just on tech Twitter.
The creator is careful to add the caveats, and I appreciate that. Many benchmarks are saturated. Anthropic has accused Moonshot of distilling from their models, meaning training on outputs scraped from another lab’s system. That claim is unresolved. And general purpose performance across all tasks still favors the big closed models. Kimi K3 wins in specific lanes, not everywhere.
🧠 The part worth thinking about
His structural argument is the most useful takeaway. Open labs ship the moment a model finishes training. Closed labs sit on models for months running safety evals and post training. Anthropic reportedly had a model internally for five months before release. So the gap between open and closed isn’t what the leaderboard shows today, it’s probably closer to 8 to 10 months of hidden lead.
He also lays out who wins when open models get good: inference providers, application builders, chip makers, and anyone shipping products on top. Models get cheaper, more tokens get used, better apps get built. The closed labs are the ones feeling the squeeze.
The risk he names is worth holding onto. If US companies build on Chinese open models, and those models get optimized for Chinese chips, that creates a real dependency. Not a moral problem, a supply chain one.
What to actually do with this
If you’re deciding on a model right now:
- Test Kimi K3 for front-end work, web engineering, and writing tasks specifically
- Measure cost per finished task, not cost per token
- Assume it’ll be slower than what you’re used to, and check whether that matters for your use case
- Keep a closed model in the mix for broad general reasoning
The full video has the live Rubik’s Cube test, the benchmark charts, and his read on the US versus China race. Worth watching the whole thing.
What’s your take, is the open weights gap actually closing?