Mira Murati's Thinking Machines Lab makes Tinker, its API for fine-tuning language models, generally available, adds support for Kimi K2 Thinking, and more
Tinker is a dream for multi-agent setups, Nathan Lambert / @natolambert : Please add olmo3 @johnschulman2 et al. The goal is to make it the foundational research infrastructure for academic LLM work. Would love to have it be seamless with Tinker! Rowan Zellers / @rown : Today we are releasing Tinker to everyone, and now with vision input! You can now finetune a frontier Qwen3-VL-235B on your own image+text data, bringing your own algorithm (sft, RL, something else?). We'll take care of the GPU infra. Full update: https://thinkingmachines.ai/ ... Devendra Chaplot / @dchaplot : Tinker is now open to everyone! We are also adding: - Vision support with Qwen3-VL - New model: Kimi K2 Thinking (1T params) - OpenAI API-compatible inference Start training models within minutes: https://thinkingmachines.ai/ ... Mira Murati / @miramurati : We're making Tinker generally available with new models and features. @thinkymachines : Tinker is now generally available. We also added support for advanced vision input models, Kimi K2 Thinking, and a simpler way to sample from models. https://thinkingmachines.ai/ ... LinkedIn: Devendra Chaplot : Tinker is now open to everyone! — We are also adding: — Vision support with Qwen3-VL — New model: Kimi K2 Thinking (1T params) … Bluesky: Nate / @zzstoatzz.io : i made a little CLI to make this easy to try out the new Tinker API — you just need — uv — a Tinker API key (tinker-console.thinkingmachines.ai/ keys) — TINKER_API_KEY=<your-key> uvx —from git+https://github.com/zzstoatzz/ hello-tinker tinker-chat [embedded post]
Context & Ripple Effects
Tinker began as a private-beta fine-tuning API for Qwen and Llama; its initial private-beta launch positioned Thinking Machines Lab around training infrastructure rather than a single proprietary model. General availability broadens that initial infrastructure bet to more developers and research teams.
The expansion comes after the lab’s $2B seed financing, which gave it backing from major compute and enterprise investors. Adding reasoning-model and multimodal fine-tuning support makes the product’s scope more consequential: it is becoming a layer for adapting externally supplied models while the lab operates the underlying GPUs.
First-order effects
- Developers can use Tinker broadly to fine-tune Kimi K2 Thinking and supported vision-language models on their own data and algorithms without directly managing the training GPU stack.
- OpenAI API-compatible inference and simpler sampling reduce integration work for teams whose existing applications and tooling already target that interface.
Second-order effects
- Model providers whose weights are supported gain another route to developer adoption, while fine-tuning platforms must compete on model breadth, multimodal support, workflow compatibility, and infrastructure simplicity.
- By packaging training infrastructure with model access, Tinker can shift some experimentation spending from self-managed GPU environments toward an API-managed service.
Third-order effects
- If this pattern persists, the competitive layer in open and accessible models may move from base-model release alone toward the infrastructure that makes adaptation, evaluation, and deployment repeatable.
- The product points to a more modular AI stack in which model creators, fine-tuning platforms, and application teams can specialize separately—though that depends on sustained access to capable models and GPU capacity.
The trend: AI companies are commercializing the operational layer around model customization, turning fine-tuning and inference infrastructure into a primary route to ecosystem control.