Thinking Machines' Inkling Model Creates a New AI Service Business: Sell the Free Trial, Then Sell the Fine-Tuning Migration.

Thinking Machines just gave away a frontier-scale model for free -- and put a price tag on the one thing that actually creates lock-in.

What happened

Thinking Machines released Inkling, a mixture-of-experts model with "975B" total parameters and "41B" active parameters, pretrained on "45 trillion tokens" of text, images, audio, and video, with a context window of "up to 1M tokens." A smaller preview, Inkling-Small, ships at "276B" total and "12B" active parameters using a similar training recipe.

The published benchmark numbers are strong for an open-weight release: SWEBench Verified at 77.6%, Terminal Bench 2.1 at 63.8%, GPQA Diamond at 87.2%, AIME 2026 at 97.1%, and VoiceBench at 91.4%. The full weights are on Hugging Face. The model is already deployed through TogetherAI, Fireworks, Modal, Databricks, and Baseten, and integrated with vLLM, llama.cpp, and Hugging Face transformers -- on day one. Access to the model through the Tinker platform comes with a "50% discount for a limited time," and Playground access is free "for a limited time" as well.

The tension that creates the business

Hacker News commenters spotted the actual strategy fast. One noted the monetization plan is a "great business model" -- companies get to own a personalized, fine-tuned model while Thinking Machines sells the infrastructure (Tinker) that makes fine-tuning practical. Another pushed back: there is "no moat" in LLM services, because switching between open-weight models requires minimal effort -- "just changing an API endpoint."

Both are right, and that is exactly the gap a service business can sit in. If switching models is trivial, most engineering teams still won't do the work of proving a switch is worth it, and even fewer will build out a fine-tuning pipeline on a brand-new platform for a single project. The commodity (open weights) really does have no moat. The judgment call of whether to switch, and the execution of fine-tuning once you do, is still expensive to build in-house.

The business idea

Package three tiers around this exact gap:

1. An evaluation engagement -- run Inkling against a client's current closed-model vendor on the client's real tasks, not just the published leaderboard numbers, and report back whether the switch is worth the engineering time.

2. A migration engagement -- for clients who decide to switch, stand up their workflow on one of Inkling's day-one hosting partners (TogetherAI, Fireworks, Modal, Databricks, or Baseten) with the right context window (64K or 256K) for their use case.

3. A fine-tuning retainer -- the sticky, recurring piece. Use Tinker while its discount runs to build a client-specific fine-tuned version of Inkling-Small, then keep billing to retrain it as their data evolves. This is the part with the actual moat, because it captures a client's proprietary data and workflow, not just a model choice.

Money play

1. Sell a fixed-scope "open-weight model evaluation" that benchmarks Inkling against a client's current closed-model vendor on the client's own tasks, using the published scores as a starting point rather than a verdict.

2. Package a migration offer for teams that decide to switch: stand up the client's workflow on Inkling through one of its day-one deployment partners so they are not locked into a single host.

3. Charge for the fine-tuning layer itself using Tinker's limited-time discount, then bill an ongoing retainer for retraining as the client's data changes -- this is the recurring revenue line.

4. Use the free Playground access while it lasts as the pitch's proof step, letting prospects see live output on their own prompts before committing budget.

5. Sell speed as the pitch: since switching between open-weight models is just an endpoint change, your service is the one that captures the client's fine-tuning data and workflow before a competitor gets there first.

Bottom line

Thinking Machines open-sourced the commodity and priced the lock-in. The weights are free, the hosting is everywhere, and the discount on the one sticky layer -- fine-tuning -- is temporary. That is a clear map for a service business: sell judgment on whether to switch, execution on the migration, and a retainer on the fine-tuning that actually creates the moat the model itself doesn't have.

Sources:

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

https://news.ycombinator.com/item?id=48924912