Free weights, paid customisation: two things have to be true for the Inkling strategy to work

Bro's Insights
, ,
1

In July, Thinking Machines released Inkling, a 975B-parameter model trained on 45 trillion tokens, under Apache-2.0. The licence permits commercial use without restriction. Tomasz Tunguz of Theory Ventures read the release alongside a $24,950 pickup truck that ships without paint, and called it a new business model: put out a deliberately unremarkable general-purpose base cheaply, then charge for customisation. His line is the cleanest summary of it. The weights are free, but the customisation charges rent.

It is a tidy story. For the strategy to actually work, two things have to be true at once, and both are currently open.

TL;DR

  1. Thinking Machines released a large model under Apache-2.0 and charges on Tinker, its own fine-tuning platform. The line between free and paid now runs between general-purpose and specific-to-you.

  2. The first premise is whether the ordinariness was chosen. Forbes reads the same model as underperforming the leading Chinese open models, a fact the company acknowledges, and casts it as the permitted low-risk option filling a regulatory vacuum rather than a strategic blank canvas.

  3. The second premise is whether the customisation revenue can be captured. Apache-2.0 lets anyone fine-tune this model anywhere and sell the result. Unlike a truck, this accessories market gives the manufacturer no priority.

[1] An unpainted truck and a grey model

Start with the analogy Tunguz uses. In June, Slate Auto revealed a $24,950 electric pickup with hand-crank windows, no stereo, no speakers, no touchscreen and no paint. The vehicle itself leaves little margin, but it is a base that invites additions, and the additions are where value accrues.

A month later came Inkling. In Tunguz’s description, if models were colours this one would be grey: a generalist, useful across many domains and exceptional in none. The release landed on Tinker, the company’s fine-tuning platform. Take the model for nothing; pay from the moment you shape it to your own use.

Another line in a familiar sequence, drawn somewhere new. Vercel drew it between framework and control, Microsoft between kernel and infrastructure, Ollama along where the workload runs. This one runs between general and specific. Free while it is useful to everybody, priced on the way to being useful only to you.

One disclosure belongs here. Tunguz is not a neutral commentator on this thesis; he has money on it. He led Ollama’s $65M round last month, and the proposition that models commoditise while the layer that runs and shapes them holds value is his portfolio position. The piece is analysis and an explanation of a bet at the same time, and it reads more accurately with that in view.

[2] First premise: was the ordinariness chosen?

Calling something a strategy carries a condition. You had to be able to do better and decline. A company that could build a sports car and ships a bare truck is executing a strategy. A company that can only build trucks, saying the same words, is narrating a constraint.

Forbes reads it as the latter. It notes that Inkling underperforms the leading Chinese open models and that the company acknowledges as much, that the architecture follows DeepSeek-V3, and that training leaned on synthetic data from Chinese models such as Kimi K2.5. From there it recasts the model’s position: not the best base model, but the permitted, low-risk one. As American firms that touch government work or fear regulatory exposure find Chinese open models increasingly untouchable, something has to fill that space, and Inkling fills it.

Two readings of one set of facts. Ordinariness as design or ordinariness as ceiling. Free release as demand generation or as the sensible move in a market you cannot win outright. Nothing published so far settles it.

[3] Second premise: can the customisation be held?

The second one matters more. Grant that the ordinariness was designed. Charging rent on customisation still requires that the customisation come through Thinking Machines, and the licence works against exactly that.

Apache-2.0 permits commercial use with no copyleft. Anyone can take Inkling, run it on any cloud, fine-tune it with a competing platform and sell the result. Specialist providers, the large clouds, an in-house ML team: all hold the same rights. The moment the weights went out, the entire customisation market opened, and Tinker became one option inside it.

This is where the truck analogy leaks. Truck accessories are physical parts, and a manufacturer retains some grip on specification and distribution. Model customisation offers no equivalent grip, because the licence already granted everyone identical rights. So if Tinker does collect rent, it will not be because the structure guaranteed it, but because the team executed better than other fine-tuning providers. Advantage from position and advantage from execution are different things, and only one of them holds when someone else tries harder.

[4] The invoice settles it

Which leaves a single test. Does money actually reach Tinker?

If the strategy reading holds, the free model generates customisation demand and a meaningful share of that demand routes to the platform built by the team that made the model. If the regulatory-vacuum reading holds, the picture differs: enterprises adopt Inkling because it is the permitted option, then shape it with the cloud contracts and tooling they already have. The model spreads and the rent goes uncollected.

Twelve months should make the split visible. If Inkling deployments climb while paid Tinker usage does not follow, this is closer to a constraint narrated well than to a business model. If both climb together, then a team held the customisation layer despite an unrestricted licence, and open-source commercialisation gains a reference case worth studying.

[5] What the case leaves behind either way

Whatever the two premises resolve to, the position of the line is worth recording. Commercial boundaries in open source have mostly run between layers: free tool with paid control above it, free SDK with paid hosting above it. This one runs along the state of a single object. General-purpose is free; fitted to your purpose costs money.

That is a different cut from the ones software usually makes. Not feature count, not support, not where it runs. How closely it has been shaped to you. As models commoditise faster, the same cut may appear elsewhere, and this is the first case to reason from.

What to check, by role

If you are evaluating open models, price the weights and the customisation separately. A free licence sets none of your real cost. What your use case requires, and where that work happens, sets it.

If you are designing an open-source business, read what your licence permits in the revenue segment before you choose it. Fully permissive terms open the commercial layer above your project to everyone. Be able to say whether your position there rests on structure or on execution.

If you invest, keep the announced business model and the collected revenue in separate columns. The free release is a verifiable fact. Whether it converts into paid customisation is still an open question.

:person_raising_hand: Lenses for the next story

First, “deliberately ordinary” is a claim that needs a falsifier. Without evidence that better was available and declined, strategy and ceiling look identical from outside. Benchmarks and a company’s own admissions are the material that separates them.

Second, the licence sets the competitive intensity of the revenue segment. The more permissive the terms, the more crowded the commercial layer above, and the less priority the original team retains. Read what was released under which licence, and the revenue path becomes legible.

Third, read the position of whoever is explaining the model. When an analyst describing a business model has invested in that model, the piece is analysis and a bet at once. That is a condition to read it under, not a reason to dismiss it.

References