Inkling's Launch Reveals the Growth Play: Give Away the Part With No Moat, Charge for the Part That Creates Lock-In.

The Play

Thinking Machines released Inkling, a large mixture-of-experts model, with full weights on Hugging Face and same-day support across TogetherAI, Fireworks, Modal, Databricks, and Baseten, plus vLLM, llama.cpp, and Hugging Face transformers integration. Playground access to try the model is free "for a limited time." None of that is where the company makes money.

The actual product is Tinker, the fine-tuning platform, which launched alongside Inkling with "50% discount for a limited time." That is the layer Thinking Machines wants you inside of, and it is priced -- just discounted enough to lower the cost of a first commitment.

Why it worked

A Hacker News comment on the launch named the dynamic directly: there is "no moat" in LLM services, because switching between open-weight models is "just changing an API endpoint." Most companies would treat that as a threat to route around. Thinking Machines built the launch around accepting it.

If the model itself has no moat, fighting to lock customers into one hosting environment for it is wasted effort -- a competitor's next release makes that lock-in irrelevant anyway. So instead of restricting where Inkling runs, Thinking Machines shipped it everywhere on day one. That removes the single biggest objection a technical buyer raises during evaluation -- "what if we're stuck with this host" -- before it's even asked.

That ubiquity does double duty: it maximizes trial volume (more places to test the model means more people testing it) while funneling all of that trial traffic toward the one product that isn't commoditized. A fine-tuned model built on a customer's proprietary data can't be replicated by switching an API endpoint. That's the layer with real lock-in, and it's the layer that's priced -- discounted for launch, but priced.

The growth play to steal

1. Identify which part of your product has no real moat -- the part any competitor could replicate or any customer could swap out easily -- and give that part away on purpose rather than trying to defend it.

2. Ship that free layer everywhere a buyer might already operate, so "where can I even run this" is never a reason to delay evaluation.

3. Reserve pricing for the layer that captures something customer-specific -- their data, their fine-tuning, their workflow -- since that's the only layer a config change can't undo.

4. Use a limited-time discount rather than a hard paywall on that sticky layer at launch, so the first commitment is cheap without being free.

5. Pair a free, time-boxed sandbox with the priced product so prospects can validate output quality before spending on the thing you actually monetize.

6. When commentators point out that your core product is commoditized, let that argument stand -- it does the selling for whichever layer you do charge for.

Bottom line

Thinking Machines didn't try to out-argue the "no moat" critique of open-weight models. It agreed with it, gave the commoditized layer away on every platform at once, and put its pricing entirely on the one layer that actually creates lock-in. That's the transferable move: stop defending the part of your product a competitor can copy, and make sure giving it away drives demand for the part they can't.

Source: https://thinkingmachines.ai/news/introducing-inkling/