Enterprise Model Forking Platform

Thinking Machines is aiming at a more valuable layer than chat, it wants to become the system a company reshapes into its own model stack. The key difference is control. Instead of giving employees one shared assistant with prompts and files, it is building tooling so a bank, drug company, or defense team can tune model behavior, plug in private data, and set separate safety rules for each workflow without rebuilding everything from scratch.

• The product direction already points to managed customization, not just model access. Tinker launched as a fine tuning API for open weight models on Thinking Machines infrastructure, letting users change training algorithms and data while the company handles distributed training, scheduling, and recovery. That is much closer to model operations software than to a chat app.
  2 thinkingmachines
• OpenAI has added Custom GPTs and model choice inside ChatGPT, but that is still mainly an interface layer on top of hosted models. For deeper changes, OpenAI separates out API fine tuning and a custom models program. That split shows how hard it is for chat products to also be full model forking environments.
  3 openai 4 openai 5 openai 6 openai
• The enterprise precedent is the data platform world. Databricks is selling governed model serving, vector search, and agent infrastructure tied to company data and permissions. Thinking Machines is pushing further down into the base model itself, where customers can change not just prompts and retrieval, but the model variant and safety policy that sit underneath the application.
  7 sacra 8 databricks

If this works, the market shifts from renting one general model to running many company specific variants off a shared base. That creates higher switching costs, because the value moves into each customer’s tuned weights, safety settings, and deployment workflows. It also turns model labs into infrastructure vendors that can capture long term enterprise spend, not just seat based chat revenue.