Tinker drops its waitlist, adds Kimi K2 and vision models

The friction of fine-tuning large models has always been a mess of waitlists, custom SDKs, and clunky inference setups.

Thinking Machines just killed all of that. On December 12, the company took Tinker out of waitlist mode into general availability, while rolling out four major updates: support for the massive Kimi K2 Thinking model, OpenAI API compatibility, and vision fine-tuning powered by Qwen3-VL.

Trillion-parameter reasoning and image inputs

The headline model addition is Kimi K2 Thinking. Packing one trillion parameters, it is the largest model hosted on Tinker so far, built specifically to tackle long chains of reasoning and autonomous tool use.

On the multimodal side, Tinker introduced two vision models: Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct. Users can now feed raw images, UI screenshots, and technical diagrams directly into both supervised fine-tuning and reinforcement learning pipelines.

To show off how vision fine-tuning works, Thinking Machines released a recipe that turns visual classification into a text generation task. Instead of training a traditional classification head, the model simply looks at the image and generates the class label.

They tested Qwen3-VL-235B against a standard vision baseline, DINOv2-base, across four datasets: Caltech 101, Stanford Cars, Oxford Flowers, and Oxford Pets. In extreme low-data regimes where each class had only one labeled example, Qwen3-VL comfortably beat DINOv2, proving that built-in multimodal context gives it a massive head start over pure vision transformers.

Sampling live weights through the OpenAI API

Beyond new models, Tinker resolved one of the biggest headaches in fine-tuning workflows: how you inspect results while models are still learning.

The platform now offers an OpenAI API-compatible scaffolding. By providing a Tinker URI pointing to a specific checkpoint path, developers can use the standard OpenAI Python client to sample predictions.

Crucially, this sampling works even while the model is actively training. Any existing tool, eval pipeline, or production scaffolding that expects an OpenAI completion call can now connect directly to Tinker weights without custom client code.

Why it matters

Fine-tuning trillion-parameter reasoning models and massive vision models usually requires specialized infrastructure teams. Opening Tinker up to everyone democratizes access to state-of-the-art open weights without the ops nightmare.

Adding OpenAI API compatibility is a brilliant tactical move. Most developers already have evaluation scripts, prompts, and agent workflows written around the OpenAI endpoint standard. Being able to drop a live checkpoint URI into that existing machinery and sample while training cuts testing loops from hours down to minutes.

Treating vision classification as language generation also changes the game for scarce-data environments. If you only have a handful of labeled images, fine-tuning a vision-language model is dramatically more sample-efficient than training traditional vision backbones.

That said, high efficiency in labeling does not equal cheap inference. Running a 235B MoE vision model or a trillion-parameter reasoning model in production costs real money, so builders will need to weigh prompt latency and hosting bills against the accuracy gains.