I told a room of business owners that open models were a distraction. Eighteen months later, three of them sit one step behind the frontier and the Linux Foundation just hired the industry’s most credible analyst to shepherd them.
The Hook
Somebody asked me a question from the back of a room in early 2025, and I gave them a confident answer that I now believe was wrong.
The question was whether they should be paying attention to open-source AI models. I said what I actually believed at the time: that for a small business owner, open models were a distraction. That the frontier labs were far enough ahead that the gap would keep widening. That running your own model was a hobby for people who liked configuring things more than they liked serving customers. I remember the phrasing. I said, “You do not have a compute problem. You have a customer problem.”
I still think the second sentence was right. The rest of it has not aged well.
Here is the direct answer, honestly stated. I was wrong about the gap. Open-weight models did not fall further behind. Kimi K2.6 now matches Qwen 3.6 Max and DeepSeek V4 and sits just behind the top closed models. Commercial model routers list all three alongside GPT and Claude as first-class options. And the movement has just acquired institutional infrastructure rather than volunteer goodwill. What I got right is that most small businesses should not self-host today. What I got wrong is that they should ignore it, because the reason to pay attention was never cost. It was control.
Key Takeaways
- Leading open-weight models have closed most of the capability gap with the closed frontier, and are now offered as peer options inside commercial model routers.
- The most credible independent analyst of frontier model trajectory has moved into the role of Global CTO of AI at the Linux Foundation, which signals institutional backing rather than community goodwill.
- The real argument for open models was never price. It is control over your data, your deployment environment, and your access.
- Vendor risk in AI is not theoretical, as demonstrated by export controls that were imposed and then lifted within a matter of months.
- The right posture for most small businesses is not to migrate, but to maintain the credible ability to migrate.
The Problem
I have a rule about being wrong in public, which is that you have to do it in the same room where you were confident.
The trouble is that in the AI space, the incentives run hard the other way. Everybody is building an authority position. Authority positions punish revision. So the ecosystem is full of people who confidently declared something in 2024, watched it fail to come true, and simply started saying a different confident thing in 2026 without ever connecting the two.
I think that is corrosive, and I think entrepreneurs can feel it. It is a big part of why so many business owners have quietly stopped trusting AI commentary, mine included. If nobody ever says “that thing I told you was wrong,” then no advice can ever be checked, and if no advice can be checked, all of it is entertainment.
So here is the check. In early 2025 I believed the frontier labs would pull away. I had reasons. Capital concentration, compute access, talent gravity, and the plain observation that the open models available at the time were meaningfully worse at the tasks my audience cared about. Those reasons were not stupid. They were just insufficient, because they were reasons about capability, and capability turned out not to be the variable that mattered.
What I did not weight properly was that a technology can converge on capability while diverging wildly on the terms under which you are permitted to use it.
But what if the question was never “which model is smarter?”
The Evidence
One. The capability gap closed. Kimi K2.6 now matches Qwen 3.6 Max and DeepSeek V4 and falls only just behind the top closed models. Three separate open-weight families sit at rough parity with each other, one step from the frontier. This is not the trajectory of a technology falling further behind.
Two. The distribution changed. All three of those models now appear as selectable, first-class options inside commercial model routers, listed next to GPT and Claude. When a product manager building an application lists your model in the same dropdown as the frontier, the market has already made its judgment about whether you are a hobbyist option.
Three. The institutional backing arrived. Nathan Lambert, who has spent years as arguably the most credible independent analyst of frontier model training and trajectory, has moved from the Allen Institute for AI into the role of Global CTO of AI at the Linux Foundation. His stated goals are to provide clarity on the evolution of frontier models, to create a vibrant and diverse open model ecosystem, and to build the institutions that make those goals possible. Institutions bring governance, standards, and permanence. Communities bring energy. The open model movement now has both.
Four. Vendor risk turned out to be concrete rather than theoretical. Export controls on Claude Fable 5 and Mythos 5 were in force, and then on June 30 they were lifted, and Fable 5 became globally available the next day. Notice what that sequence demonstrates. A capability your business depends on can be geopolitically switched off and switched back on, on a timeline you do not control, for reasons that have nothing to do with you. Similarly, GPT-5.6 was restricted to a small group of trusted partners before its public release was cleared. Access is a policy variable now.
Five. And the underlying economics keep validating the direction. RAND found that more than 80 percent of AI projects fail, driven largely by fragmented data across disconnected systems and missing governance. MIT found roughly 95 percent of generative AI pilots produce no P&L impact. Both point at the same underlying constraint: the AI needs to be where your data is, under conditions you can govern. That constraint gets easier to satisfy, not harder, when the model can run inside your own walls.
The conventional narrative was that open models were the cheap option for people who could not afford the good one. The evidence says open models became the sovereign option for people who cannot afford to have someone else hold the switch.
What Changed for Me
The thing that actually moved me was not a benchmark chart. It was a conversation with a client in a regulated industry who had built an entire practice around not using AI, because her compliance obligations meant her client data could never leave her environment.
For two years I had accepted that as a hard boundary. It was the polite version of “AI is not for you.”
Then I sat down and worked through what it would actually take for her to run a capable model on her own hardware, on her own data, with no third party in the path. Not as a science project. As an operating decision, with real numbers. Hardware. Electricity. Maintenance hours. Her hourly rate. Quality measured against her actual work, not against a public benchmark.
The answer was that it was entirely feasible, and it had been feasible for at least six months, and I had not told her because I was still repeating a position I had formed in 2024.
That was the moment. Not the model quality. The realization that my advice had a shelf life and I had not been checking the date.
What I believe now is narrower and, I think, more useful. Most small businesses should not self-host today, because most small businesses do not have data that requires it and do not have the hours to maintain it. That part of my original answer stands. But every business should know what it would cost to leave, should have run the numbers once, and should hold the credible ability to move. Not because you are going to. Because being able to changes how you negotiate, how you plan, and how you sleep.
Optionality is worth more than optimization. That is the sentence I would give the person in the back of that room, if I could go back.
Practical Steps
1. Sort your data into three buckets, in writing. What you would happily send to any vendor. What needs contractual protection. What should never leave your environment. Most business owners have never done this, and the third bucket is usually smaller and more specific than they fear.
2. Ask your vendor’s terms what they are actually permitted to do with your inputs. Read the data usage section. Have an AI translate it into plain English. Then ask what your customers assume you are doing with their data. The gap between those two answers is your real risk, and it is usually a surprise.
3. Run the leaving math once. Model what it would cost to run your highest-volume workload on an open-weight model in your own environment. Hardware, power, maintenance hours at your real hourly rate. You are not doing this to migrate. You are doing it so the number is not a mystery.
4. Test an open model on your actual work, not on a benchmark. Take twenty real examples from your business. Run them through a leading open-weight model and your current closed model. Score them against what you would actually accept. This takes an afternoon and it will settle the question for you personally.
5. Move one low-risk production workload, just to prove you can. Not the important one. The boring one. The point is to discover the friction while nothing is at stake, so that if you ever need to move something important, you already know where the sharp edges are.
6. Keep the frontier model for the hardest ten percent. This is not a religious choice. Most real workflows have one genuinely hard reasoning step and several routine ones. Send the hard step to the best model available and run the rest wherever you like.
7. Write down what would make you move, before you need to. A price increase of a certain size. A terms change. An access restriction. Decide the trigger while you are calm. Vendor risk arrives on a Tuesday, and Tuesday is a bad day to be forming policy.
Frequently Asked Questions
Are open-source AI models good enough for business use in 2026?
For a large share of production workloads, yes. Leading open-weight models including Kimi K2.6, DeepSeek V4, and Qwen 3.7 now sit close behind the top closed models and appear as first-class options inside commercial routers. The remaining gap shows up mainly on the hardest reasoning tasks.
Why would a small business run its own AI model?
Control rather than cost. Running a model in your own environment means sensitive data never leaves your walls, your access cannot be restricted by a vendor or a regulator, and your pricing cannot change with thirty days notice. For most businesses this matters for a subset of data rather than for everything.
What is vendor risk in AI, practically speaking?
It is the risk that the capability your business depends on is repriced, deprecated, or made inaccessible for reasons outside your control. Export controls on frontier models were imposed and then lifted within months in 2026, and some models were restricted to selected partners before public release.
Should I switch entirely from a closed AI model to an open one?
Usually not. The stronger position is a hybrid: run high-volume routine work on an open or low-cost model, reserve the frontier model for genuinely hard reasoning steps, and maintain the tested ability to move. Optionality is more valuable than a single optimized choice.
What does the Linux Foundation’s involvement in AI mean for open models?
Institutional backing brings governance, standards, and permanence that a volunteer community cannot provide on its own. The movement of a leading independent frontier-model analyst into the Global CTO of AI role signals that open weights are becoming durable infrastructure rather than a temporary alternative.
The Close
Somebody asked me a question from the back of a room, and I answered it with more certainty than I had earned.
I have thought about that person a lot. I do not know if they took my advice. If they did, they spent eighteen months not paying attention to something that was quietly becoming one of the more important developments in their field, and they did it because a guy with a microphone sounded sure.
That is the actual cost of confident wrongness, and it does not land on the person holding the microphone.
So let me say the thing I would want said to me. The open models are real now. Not cheaper. Real. They run one step behind the frontier, they run on hardware you can own, and they are now backed by the same kind of institution that made open source software the invisible foundation of the entire internet. That transition took twenty years the last time, and it was obvious only in retrospect, and everybody who called it a hobby in year three felt very reasonable at the time.
You do not need to migrate. You do not need to become a systems administrator. You do not need to develop opinions about quantization.
You need to know what it would cost you to leave. You need to know which of your data can never go somewhere else. And you need to hold, somewhere in the back of your mind, the knowledge that the switch to your most important business capability is currently in somebody else’s hand.
I was wrong about the gap. I am telling you here, in the same room, at the same volume.
Do not let anybody, including me, hand you certainty about a field that rewrites itself every eighteen days.
About the Author
Jonathan Mast is the founder of White Beard Strategies and serves a community of over 200,000 entrepreneurs learning to use AI in their businesses. He created the Perfect Prompt Framework, speaks internationally on practical AI implementation, and tries to make a habit of correcting himself in the same room where he was wrong.
