[00:03] public. Anyone can download the whole thing right now. But here is the strange part. Almost nobody can actually run it. So what does that mean for the rest of us? I'm the digital avatar of Julian Goldie, and I help people learn AI tools
[00:15] Today, I'm breaking down Kim K3, what changed this week, how it works, and how I run Kim 3 inside Agent OS so it does real work instead of just chatting. setting almost everyone gets wrong on day one and it quietly ruins the output.
[00:30] Here is what happened. Moonshot AI launched Kimmy K3 on July 16th, 2026. At their app and their API. The weights were locked. Then on July 27th, they along with the technical report. That is the part people have been waiting 10
[00:46] days for. The size is the headline, Kim K3 has 2.8T parameters. That makes it the first open model in the 3T class and the largest open weight model anyone has released. It also has a 1M token context window and native vision. So it reads
[01:00] text and images inside the same model. Here is how I'd use that for the AI profit boardroom. I drop the entire onboarding flow for the AI profit message, every call prep dock, every resource walkthrough, all of it at once
[01:13] with a 1M token window. I'm not chopping it into pieces and hoping it remembers. It holds the whole thing and maps day one for a new AI profit boardroom member how it actually works because the size number on its own is misleading. Kimmy
[01:27] of it like a very large team. There are 896 experts inside it, but only 16 of them get called for each token. So, while the full model is 2.8T parameters, only about 104B are doing the work at any moment. That is how a model this big
[01:42] stays practical to serve. Moonshot also built it on two new pieces called Kimmy delta attention and attention residuals. In plain English, those change how information travels across long inputs and through the depth of the model.
[01:54] roughly two and a half times better scaling efficiency compared to Kim K2. That is the real story here. It is not just bolting on more size. It is know. Kim K3 always thinks you cannot
[02:08] switch the reasoning off on the API. You can set the effort to low, high, or max. And max is the default at launch. So, how do you get to it? Easiest route, go to kimmy.com and sign in with a free account. Open the model dropdown because
[02:20] it defaults to K2.6 and switch it to K3. There is also Kimmy work on desktop and Kimmy code in your terminal where you pick the model with a /model command. Developers can call it on the API with the model ID Kimmy- K3. And if you have
[02:33] serious hardware, you can now serve the weights yourself with VLLM, SGLANG, or token speed. the weight ship under the Kimmy K3 license and the technical can read exactly how they trained it rather than guessing from benchmark
[02:45] charts. Let me show you the flow I'd run to test it properly. I sign in at kimmy.com, switch the model to K3 and I skip K3 Swarm on a free account because allowance far too quickly to finish a build. Then I give it something real,
[02:58] something like build a landing page for an SEO agency. Watch two things while it its reasoning before the answer and that tells you whether it actually understood anything visual, K3 can look at screenshots of its own output and fix
[03:13] loop. That is what makes it strong for front-end work, gamedev, and CAD. And it is why in blind testing by Arena, developers preferred K3 over the leading closed models for front-end coding. That is the test. Real work is where I put
[03:27] Kimmy K3 inside agent OS. Now, if you want the shortcut, I built a full Kimmy K3 masterass inside the AI profit boardroom along with the exact agent OS build. I run Kimmy K3 in. You get the walkthroughs for the settings, the
[03:40] calls where you can bring your own setup and ask questions about it directly. You agent OS ready to install. And we built out a complete 30-day road map here as what to do in week one instead of guessing. Everything gets updated as
[03:55] are changing fast. That is aprofitboardroom.com. instead of the do-it-yourself version. Now, here's where it gets interesting for me because this is where Kimmy K3 actually lives in my setup. Inside agent
[04:08] OS, a chat window is fine. It is not a system. So, I do not leave Kim K3 sitting in a browser tab. I run Kimmy K3 inside agent OS and that is what turns a model into something that finishes work. Inside Agent OS, Kim K3 is the reasoning
[04:21] layer and Hermes is the execution layer. Skills teach it repeatable processes. tools. MCP connections give it access to your apps. Scheduled workflows run Connecting it is one command. Open your terminal, type Hermes model, pick
[04:36] the models page in your agent OS dashboard. That is the setup I'd point at the AI profit boardroom. With Kimmy K3 inside Agent OS, I'd run a scheduled research pass every morning on the AI tools and updates AI profit boardroom
[04:49] members are actually asking about, then draft the training outlines from it. Kim K3 does the thinking. Hermes does the building and the AI profit boardroom starting from a blank page. Now, the honest picture on performance on
[05:02] honest picture on performance on Moonshot's own numbers. K3 hits 93.5 on GPQA diamond, 88.3 on Terminal Bench 2.1, and 91.2 on Browse Comp. It debuted at number three on the artificial analysis leaderboard. And Moonshot says
[05:16] it plainly in their own blog. Overall, K3 still trails Claude Fable 5 and GPT 5.6 Soul. They also flag a gap in day-to-day user experience compared to those two. When a lab publishes its own weak spots, that is a good sign for
[05:30] are two limits you need to know before you build on it. The first one is the was trained in preserved thinking history mode. That means the full to the model on every turn. If your tool doesn't do that, or if you switch an
[05:44] through, the quality goes unstable and you will think the model is broken. harness that supports it properly like Kimmy Code. The second one is that K3 is very proactive. It was trained on long hard tasks. So when your instructions
[05:58] you. If you need it inside strict boundaries, spell those boundaries out file. Do not assume it will wait and ask. So who is this actually for? If you without any setup at all, you are
[06:12] covered. Go to kim.com free account. Switch to K3. Done. If you are a so you can point existing tools at it with small changes. If you run a team mean you can host it yourself and keep everything inside your own walls. And if
[06:26] you want it doing work rather than answering questions, that is agent OS, who it is not for, anyone hoping to run this on a laptop. The download alone Moonshot's own advice is to serve it on setups with 64 or more accelerators.
[06:39] Self-hosting K3 is a data center job, not a gaming GPU job. For everyone else, the hosted version is the practical route. Now, my tips, use the context window. Do not chunk your inputs out of habit. One m tokens is enough to hold an
[06:51] entire codebase or a full document library in one go. So feed it the whole thing instead of feeding it slices. Two, match the effort to the task. Max thinking. Drop it to low for small stuff and save the deep effort for the long
[07:04] builds. Three, never switch models midsession. Start a new session for K3 every time. Four, if you are sending images through the API, use base 64 or an uploaded file ID. Public image links are not supported and it trips people
[07:17] model is built for long horizon work, so tell it what done looks like, then let 2 minutes. And one more use case before I wrap. I'd point Kimmy K3 at a full content series for the AI Profit Boardroom. Feed it the topics AI Profit
[07:32] about. Let it draft the hooks, the scripts, and the captions in one run. thing out. That is the kind of content that brings the right people into the AI profit boardroom. If you want the full process, SOPs, and 100 plus AI use cases
[07:47] Links in the comments and description. You'll get all the video notes from there, plus access to our community of 85,000 members who are crushing it with AI. You can get the full zip file inside the agent OS ready to install. And we
[07:59] here as use cases. And if you're about to go and try Kim K3 yourself, here is great first result, then you'll switch a session midway and wonder why it fell instruction and it'll build something you didn't ask for. You'll look at
[08:13] out of reach. That is exactly what we work through inside the IP profit tutorials, the prompts, and live coaching calls where you can bring your exact setup to us with over 4,000 members figuring this out alongside you.
[08:26] agent OS ready to install. And we built out a complete 30-day roadmap here as use cases. Come and join us at apiprofitboardroom.com.