Author: Lin Qiao; Compiler: Peggy, BlockBeats

The shutdown of Mythos this week has made many AI entrepreneurs re-aware of an issue obscured by cost discussions: When the core capabilities of a product are built on external models and platforms, what exactly does the company really own?

In the past few years, open source models have often been discussed within the framework of "cheaper alternatives to cutting-edge models." However, this article believes that cost is not the most critical variable, control is. For an AI company, calling cutting-edge model APIs can quickly launch products and lower technical barriers, but it also means that core capabilities may be subject to the model supplier's rules, prices, strategy adjustments, and even delisting decisions.

The article further proposes that "having intelligence" does not mean giving up cutting-edge models, but that enterprises should integrate their own data, workflow, domain knowledge, evaluation standards and edge cases into a controllable model system. In the future, AI competition will not necessarily be dominated by a single largest model, but will emerge with multiple "frontiers": general frontier models, enterprise-specific post-training models, vertical dedicated models, and routing systems composed of multiple models.

The closure of Mythos is therefore like a reminder: the real moat in the AI era is not just how powerful the model can be used, but whether intelligence can be turned into the company's own assets.

The following is the original text:

Mythos was shut down this week. Whether you agree with this decision or not, it is actually not the point.

What really stings a lot of people is this: a company built on intelligence beyond its control is suddenly exposed to a set of decisions that it has no influence over. When many founders see this scene, they will ask themselves the same question: Which parts of my business are actually just "rented"?

In the past few years, much of the discussion about open source models has revolved around cost: Do they really get the job done? If so, how much cheaper would it be compared to calling the cutting-edge model API?

Now, we have a pretty clear answer. We have worked with companies such as @RampLabs, @cursor_ai, @harvey, etc., and the basic path is similar: starting from a powerful open source model, post-training with work content that is really important to the company, and continuing to use rigorous evaluation to compare it with cutting-edge models.

The results are surprising again and again. On the tasks that enterprises care about most, a tuned open source model can often approach or even reach the quality of cutting-edge models at a very low cost.

But what really became clear this week is this: cost has never been the most important issue.

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

Many discussions recently have been summarized as the difference between "renting" and "owning." The analogy isn't perfect, but it's useful.

Rent smart

The rental thing worked great until something went wrong. The apartment is ready for move-in, the lights are working, the water pipes are working, and maintenance is taken care of. That’s why most companies choose this path in the first place.

The Frontier Model API is a great product. They allow startups to build things that would have seemed unthinkable just a few years ago.

But renting also means restrictions. The landlord can raise the rent, decide what improvements you can make, and change the rules. Occasionally, for reasons that have nothing to do with you, they may tell you that it's time to move.

You did nothing wrong. You're just always operating on someone else's turf.

This is why the story of Mythos resonates with so many people. When your core capabilities are entirely dependent on someone else's platform, you are exposed to a set of decisions that are outside of your control.

Most of the time, it doesn't matter. But sometimes, it becomes extremely important in a split second.

Have intelligence

The lesson here is not that companies should stop using cutting-edge models. Far from it. Cutting edge model labs have made extraordinary technology. Most products should use them. We use it ourselves too.

In many senses, cutting-edge models are becoming infrastructure. But infrastructure and ownership are two different things.

You can use public infrastructure while still owning something that truly creates value for your business. In the field of AI, “owning” means starting from a state-of-the-art open source model and shaping it around the most unique parts of your company.

Your data.

Your workflow.

Your domain knowledge.

Your edge case.

Your evaluation criteria.

Your definition of "good".

Over time, this model will become less general and more reflective of what your company actually handles every day. It is here that value is created.

Think of it like a house. Moving furniture is easy, as is painting a wall. But if your future depends on the layout of your house, sooner or later you're going to wish you had the ability to move walls. The same goes for intelligence.

When intelligence truly belongs to you, no one can quietly take away the floor from under your product’s feet.

This is why we built Fireworks the way we did.

We put training and inference in the same system, allowing companies to adopt the best open source models, shape them around the most important problems in their business, and deploy them stably into production environments.

Not just consumer intelligence. But have intelligence.

There is no unique frontier

Another optimistic revelation this week: The future of AI does not depend on one model winning them all.

There is no single frontier. There are many kinds of frontiers.

The frontier model is a frontier.

A model post-trained on years of company proprietary knowledge is another frontier.

A specialized model that solves a narrow problem better than any other model is another frontier.

A system that can route requests to multiple models and let them work together to perform many tasks beyond a single model is also cutting-edge.

The most interesting change in the field of AI is not that a certain model is getting smarter, but that intelligence is becoming more and more customizable.

The companies that ultimately win are not necessarily those with the biggest models, but those that can turn intelligence into their own unique assets.

Looking to the future

A lot of time this week has been spent reacting to the news, and we’ve chosen to continue releasing products: @Kimi_Moonshot K2.7 Code, @MiniMax_AI M3, @Alibaba_Qwen 3.7 Plus.

The future I look forward to is not one where a model quietly devours everything it sees.

But many teams can have their own part of the frontier.

If Mythos being shut down has you rethinking your trade-offs, we'd love to chat.