Author: Lin Qiao

Translation: Deep Tide TechFlow

Deep Tide Introduction: Mythos was suddenly shut down this week, directly exposing a fatal risk that most founders overlook: when your core capabilities entirely depend on someone else’s platform, your survival is no longer in your hands. Who truly owns the intelligence your product relies on?

Mythos was shut down this week. Whether you agree with the decision or not, that’s hardly the point.

A company built on intelligence it cannot control suddenly finds itself exposed to decisions it cannot influence. Many founders seeing this event have asked themselves the same question: which parts of my business are actually just rented?

For the past few years, discussions about open-source models have primarily revolved around cost. Can they really get the job done? If they can, how much cheaper are they compared to calling cutting-edge APIs?

Now we have a fairly clear answer. We work with companies like Ramp, Cursor, and Harvey, using the same basic approach: starting with a powerful open-source model, fine-tuning for the truly important work in the business, and rigorously evaluating it against state-of-the-art models.

The results have been consistently surprising. On the tasks they care most about, fine-tuned open-source models can achieve the quality of cutting-edge models at a very low cost. What happened this week made one thing clear: cost has never been the most important issue.

The deeper issue is control. Who owns the intelligence that your product relies on?

Much recent discussion has been framed as renting vs owning. It’s not a perfect analogy, but it’s useful.

Renting Intelligence

Renting works well until something goes wrong. Apartments are move-in ready. Lights turn on. Pipes are clear. Someone is responsible for maintenance. That’s why most companies start from here.

Cutting-edge APIs are incredible products. They enable startups to build things that seemed impossible a few years ago.

But renting has limitations. Landlords can raise the rent. They can decide what modifications you can make. They can change the rules. Occasionally, for reasons unrelated to you, they will tell you to move out.

You haven’t done anything wrong. You’re just operating on someone else’s turf. That’s why the story of Mythos resonated with so many. When your core capabilities entirely depend on someone else’s platform, you expose yourself to decisions beyond your control.

Most of the time it doesn’t matter. Sometimes it suddenly becomes very important.

Owning Intelligence

The lesson is not that companies should stop using cutting-edge models. On the contrary. Cutting-edge labs have built extraordinary technology. Most products should use it. We do too. In many ways, cutting-edge models are becoming infrastructure. But infrastructure and ownership are two different things.

You can use public infrastructure while still owning something that creates value for your business. In the field of AI, ownership means starting with the state-of-the-art open-source models and shaping them around the uniqueness of your company.

Your data.

Your workflows.

Your domain expertise.

Your edge cases.

Your evaluation criteria.

Your definition of "good".

Over time, the model becomes less generic, more reflective of the work your company does every day. That’s where value is created.

Think of a house. Moving furniture is easy. Painting walls is easy. But if your future depends on the layout itself, in the end, you will want the ability to move the walls. Intelligence is the same.

When the intelligence belongs to you, no one can quietly pull out the foundation of your product.

That’s why we are building Fireworks this way.

Training and inference under one roof so that companies can adopt the best open-source models, shape them around the issues that matter most to the business, and reliably deploy them in production.

Not just consuming intelligence. Owning it.

There is no single frontier

An optimistic conclusion this week is that the future of AI does not depend on the victory of a single model.

There is no single frontier. There are many frontiers.

Cutting-edge models are one frontier.

Models fine-tuned based on years of proprietary company knowledge are another.

Specialized models that solve a narrow problem better are another.

Routers that map requests to a set of model collections, jointly outperforming any single model on many tasks, are also one.

The most interesting thing in the field of AI is not that a single model becomes smarter. It is that intelligence is becoming increasingly customizable. The winning companies will not necessarily be those that own the largest models. But those that turn intelligence into a unique proprietary asset.

Looking Ahead

While everyone was reacting to the news this week, we were busy releasing products - Kimi Moonshot K2.7 Code, MiniMax M3, Alibaba Qwen 3.7 Plus.

The future I look forward to is not one where a model quietly consumes everything it sees. But one where many teams possess the part of the frontier that matters to them.

If the shutdown of Mythos has caused you to rethink this trade-off, we’d be happy to chat.