Kimi K3、えぐい。

海外の個人開発者が、メインの開発環境をClaude Codeから移しました。

移した先は、ターミナルで動くコーディングAI「OpenCode」に、Moonshot AIの新モデル「Kimi K3」をつないだ構成。

・決め手は性能ではなく「5分おきに作業を止められる」という不満
・Kimiに月19ドル（約3,100円）払ってAPIキーを取得
・K3は2.8兆パラメータ、100万トークン級（このプランでは25.6万まで）

など、手を止められない環境に組み替えた話です👇

seven weeks. that's all it took.

- BrowseComp: 84.3
- DeepSWE: 59.0
- GPQA Diamond: 91.0
- FrontierSWE: 66.7
- HLE with tools: 57.9

these are not Kimi K3's numbers. these are Claude Opus 4.8's scores at max effort.

when Opus 4.8 shipped on May 28, it was the most capable model available to the public. if someone had said back then that seven weeks later, anyone would be able to download the weights of something even more capable onto their own disk, nobody would have believed it.

that's exactly what happened yesterday. Moonshot AI released Kimi K3, and it beats Opus 4.8 on four of those five benchmarks.

and this is not just an Opus story. K3 matches Claude Fable 5 and GPT-5.6 Sol on several benchmarks and beats them on some, while its API is priced at roughly a third of Fable 5.

K3 is open weights. Moonshot says it will ship all 2.8 trillion parameters on 27th July, which would make it the largest open model released so far.

why this matters?

every closed frontier model is rented intelligence, accessed through an API the provider can reprice, deprecate, or shut off.

last month proved this is not a theoretical risk. a US export directive forced Anthropic to take Fable 5 offline globally for 18 days, and teams built on the strongest model in the world simply had to wait.

weights on a disk don't have that failure mode. a downloaded model runs on private infrastructure, fine-tunes on private data, and cannot be switched off by anyone.

the benchmark results against the current frontier is on par, and that's the interesting part. on hard web research, K3 has the best score in this lineup, ahead of both Fable 5 and Sol.

on open-ended engineering, it beats Sol by ten points, though Fable 5 is still ahead there. on repo-level bug fixing and PhD-level knowledge, the closed models win, and Moonshot says so openly.

earlier, open models were expected to trail the frontier by six to twelve months. now the difference comes down to which benchmark is being compared.

it's worth remembering the weights aren't public yet, so until July 27 this is an API model like any other. but if Moonshot delivers, this level of intelligence becomes something anyone can download and own.

it's a great time to be alive.

🚀搞定！WorkBuddy + Kimi K3 配置极简教程：

1. ~/.workbuddy 目录下，新建 models.json

2. 将下面内容拷贝到json中，记得替换apikey

{
"id": "k3",
"name": "k3",
"vendor": "Custom",
"url": "https://api.kimi.com/coding/v1",
"apiKey": "你的API KEY",
"supportsToolCall": true,
"supportsImages": true,
"supportsReasoning": true,
"useCustomProtocol": false
}

3. 重启 WorkBuddy

收工！✌🏻