Thinking Machines Lab releases Tinker as a general product: Includes Qwen3 VL and the Kimi K2 Thinking Input.

Thinking Machines Lab moved to a larger location Tinker training API The release of Tinker into general availability added three major features: support for the KimiK2 Thinking reasoning models, OpenAI-compatible sampling and image input via Qwen3VL vision languages. Tinker is now a useful tool for AI engineers to refine frontier models, without having to build centralized training infrastructure.

What Does Tinker Do?

Tinker, a language training API, focuses primarily on fine-tuning large models and conceals the hard work of distributed training. Write a Python loop to run on an only CPU machine. You specify the data, RL or training environment. Tinker maps this loop to a cluster GPUs, and performs your specified computation.

API reveals a few primitives like forward_backward Calculate gradients optim_step Update weights Sample The functions to save and load the state and generate outputs. It is a way to keep the logic of training explicit and accessible to people who do not wish to deal with GPU failures or scheduling, yet want to use supervised learning, reinforced learning or preference optimization.

Tinker relies on low rank adaption, or LoRA, instead of full fine-tuning for all supported models. LoRA is a technique that uses a small number of adapter matrices to train on frozen weights. It reduces the memory requirements and allows repeated testing of dozens of expert models within a cluster.

Think about Kimi k2 thinking and general availability

Tinker’s December 2025 Update is notable for the fact that it no longer maintains a waiting list. Sign up to see current models and prices, as well as run examples of cookbooks.

Users can fine-tune the models on their side. moonshotai/Kimi-K2-Thinking Tinker. Kimi K2 thinking is a model of reasoning with approximately 1 trillion parameters total in an expert architecture. This model is intended for heavy tool usage and long chains of thinking. It’s currently the biggest in Tinker.

Kimi K2 thinking is a model in the Tinker lineup. It appears alongside models such as Llama-3, DeepSeek V3.1, Qwen3 and its mixture of expert variants. The reasoning model always produces internal chain of thoughts before the visual answer. Instruction models, on the other hand, focus more on latency.

OpenAI-compatible sampling during training

Tinker had an already native interface for sampling through its Clients can be contacted for a sampling.. This pattern is used to build a ModelInput From tokens to passes SamplingParamsCalls and. Sample Get a future which resolves outputs

This new version adds another path which mirrors OpenAI’s completion interface. Tinker allows you to reference a model checkpoint by using URIs like this:

response = openai_client.completions.create(
    model="tinker://0034d8c9-0a88-52a9-b2b7-bce7cb1e6fef:train:0/sampler_weights/000080",
    prompt="The capital of France is",
    max_tokens=20,
    temperature=0.0,
    stop=["n"],
)

Visual Input with Qwen3 VL on Tinker

Tinker now exposes 2 Qwen3-VL vision language models. Tinker now exposes Qwen3 VL language vision models. Qwen/Qwen3-VL-30B-A3B-Instruct The following are some examples of how to get started: Qwen/Qwen3-VL-235B-A22B-Instruct. These models are also available as Vision MoE in the Tinker lineup and can be used for both training and sample through the same API surface.

You can send an image to a model by creating a ModelInput There are many ways to interlace an ImageChunk With text chunks. Text chunks are available. research blog The following is a simple example.

model_input = tinker.ModelInput(chunks=[
    tinker.types.ImageChunk(data=image_data, format="png"),
    tinker.types.EncodedTextChunk(tokens=tokenizer.encode("What is this?")),
])

The following is a list of the most recent and relevant articles. image_data Raw bytes is a raw data format. You can also download the format of this page. This identifies a particular encoding. Png The following are some examples of how to use The jpeg format is a good alternative to the jpg.. The same representation can be used for supervised and RL-fine tuning. This keeps the multimodal pipelines uniform at API level. Tinker’s LoRA setup fully supports vision inputs.

Qwen3-VL Versus DINOv2 On Image Classification

Tinker Team fine-tuned the vision to show off what it can do Qwen3-VL-235B-A22B-Instruct As an image classifier. They used four standard datasets.

• Caltech 101
• Stanford Cars
• Oxford Flowers
• Oxford Pets

Qwen3 is a visual language model. Classification is therefore a text-based process. It receives an input image, and then generates a class name in text format.

In order to get a good baseline they refined a DINOv2 model. DINOv2 encodes images and embeddings into a vision transform that is used to perform vision tasks. This experiment uses a DINOv2 with a classification module to determine a distribution for the N labels of each dataset.

Both Qwen3-VL-235B-A22B-Instruct and DINOv2 base are trained using LoRA adapters within Tinker. Data efficiency is the focus. The experiment starts with only one sample for each class and increases. Each setting is measured for classification accuracy.

The Key Takeaways

1. Tinker, which is generally available now, allows anyone to sign up for a Python-based training loop and tune their open weight LLMs while Tinker takes care of the backend distributed training.
2. Platform supports Kimi K2 Thinking from Moonshot AI. It is a mixture of 1 trillion parameters experts reasoning models.
3. Tinker has an OpenAI inference tool that allows you to sample in-training checkpoints by using a tinker://… OpenAI clients and tools that are standard style can be used to model URI.
4. Qwen3 models Qwen3VL 30B and Qwen3VL 235B allow developers to build multimodal pipelines that include both audiovisual input. ImageChunk Text input using the LoRA-based API.
5. Thinking Machines shows that Qwen3VL235B fine-tuned on Tinker achieves better few shot classification performance on datasets like Caltech 101 and Stanford Cars than the DINOv2 baseline, which highlights data efficiency for large vision language model.

Asif Razzaq serves as the CEO at Marktechpost Media Inc. As an entrepreneur, Asif has a passion for harnessing Artificial Intelligence to benefit society. Marktechpost is his latest venture, a media platform that focuses on Artificial Intelligence. It is known for providing in-depth news coverage about machine learning, deep learning, and other topics. The content is technically accurate and easy to understand by an audience of all backgrounds. This platform has over 2,000,000 monthly views which shows its popularity.