By Michael Domanic, Section Head of AI

Moonshot AI released Kimi K3 last week - a 2.8 trillion parameter open-weight model from China that outperforms Claude Opus 4.8 and GPT-5.5 on most benchmarks while trailing only Claude Fable 5 and GPT-5.6 Sol. It's the largest open-weight model ever released, it has a million-token context window, and API pricing starts at $0.30 per million cache-hit input - a fraction of what the frontier US models cost. 

(Disclaimer: A top White House official is claiming that Moonshot AI secretly trained Kimi K3 by conducting large-scale, covert "distillation" - systematically querying Anthropic's Fable model to extract its behavior and use those outputs as training data.)

I'm not going to pretend I have a fully formed point of view on what this means for enterprise AI strategy. The honest truth is I'm still working through it. But here are the questions I'm sitting with right now as a Head of AI, because I think they're the right ones to be asking.

Does your whole organization need frontier models?

This is the question K3 forces you to confront. Most knowledge workers are using AI for tasks that don't require the most advanced model available - drafting emails, summarizing documents, basic analysis, straightforward Q&A. They're driving a Lamborghini to the grocery store.

K3 and models like it are increasingly capable enough for that kind of work at dramatically lower cost. 

At our AI Leadership summit this week, I interviewed the CIO of Cloudflare about the AI transformation happening in their business. They've built a custom harness where they control which models their employees access, routing simpler tasks to cheaper models and reserving frontier capabilities for the work that actually needs them.

The counterargument - and it's a real one - is that model routing isn't free. The user experience might be seamless, but the organization still needs the infrastructure maturity to deploy, govern, and maintain a routing layer. For most organizations, that's premature. But the cost math is getting hard to ignore.

The China question is real and unresolved

K3 is open-weight, which means you can run it locally - and local deployment theoretically gives you more security and control. 

But there’s a trust question involved: Can you be confident there's no backdoor? Open-weight doesn't mean open-source in the traditional sense. You can inspect the weights, but that doesn't make auditing a 2.8 trillion parameter model trivial. For regulated industries or companies handling sensitive data, that uncertainty is a real barrier regardless of what the benchmarks say.

Then there's the geopolitical risk. If the current administration decides to crack down on enterprise use of Chinese AI models, there's no advance notice and no grace period. If you've built critical workflows on K3, you'd be scrambling to migrate overnight. That's a real operational risk that the benchmarks don't capture.

The deeper problem: most companies aren't ready for this conversation

Here's what I keep coming back to. The conversation about which model to use, how to route between models, and how to optimize inference cost assumes a level of organizational AI maturity that most companies don't have yet.

I was talking to a CIO recently who was excited about deploying open-source models to cut costs. But when I asked about their workforce enablement program, they didn't really have one. Their employees were barely using the AI tools they already had access to. Optimizing model selection for a workforce that hasn't figured out how to use AI in the first place is like debating which racing tires to put on a car that nobody knows how to drive.

First, get your people genuinely capable with AI - using it daily, building with it, understanding what it can and can't do. Then, once you have that foundation, you can start getting prescriptive about which models to use for which tasks, how to manage inference costs at a granular level, and whether open-weight models from China or anywhere else have a role in your stack.

What I’m telling Heads of AI about this today

If you're an enterprise organization that's still early in AI transformation - and most are - K3 doesn't change your priorities. Focus on enablement, pick one primary platform, go deep, and build the organizational muscle. The model landscape will keep shifting, and being locked into any single model matters less than having a workforce that can adapt when it does.

If you're a more technically sophisticated organization that's already built a custom harness and has strong AI maturity across the workforce, then K3 may be worth evaluating seriously for specific use cases where you don't need frontier capabilities and the cost savings are meaningful. Just go in with eyes open on the trust and geopolitical dimensions.

And if you're somewhere in between - which is probably most of you reading this - start thinking about what your model strategy will look like in 12 months, even if you're not ready to act on it today. The days of "everyone gets Claude and that's our AI strategy" are probably numbered. The economics of inference are changing, the model landscape is fragmenting, and the organizations that think ahead about this will have more options when it matters.

See you next week, 

Michael
Your fellow Head of AI

‍