Tinker: General Availability and Vision Input

2025-12-12 · Source: Thinking Machines Lab - Connectionism · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Computer Vision · Depth: Intermediate, quick

Summary

Thinking Machines announced four significant updates to its Tinker platform on December 12, 2025, making it generally available without a waitlist. The platform now features Kimi K2 Thinking, a new trillion-parameter reasoning model designed for complex reasoning and tool use, which users can fine-tune. Tinker also introduced an OpenAI API-compatible inference interface, allowing for plug-and-play integration with other compatible platforms and enabling sampling from models even during training. Additionally, Tinker now supports vision input through the integration of Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct models, facilitating the processing of images, screenshots, and diagrams. A new cookbook recipe demonstrates fine-tuning VLMs as image classifiers, with Qwen3-VL-235B-A22B-Instruct showing superior performance over DINOv2-base in limited-data classification tasks across datasets like Caltech 101 and Stanford Cars.

Key takeaway

For AI Architects and Computer Vision Engineers building multimodal applications, Tinker's new vision capabilities with Qwen3-VL models offer a compelling solution for image classification, especially in data-scarce environments. Your teams can leverage the OpenAI API compatibility for easier integration and explore fine-tuning the Kimi K2 Thinking model for advanced reasoning tasks, potentially accelerating development cycles and improving model performance on specialized datasets.

Key insights

Tinker's general availability, new reasoning model, OpenAI API compatibility, and vision input expand its utility for AI development.

Principles

• VLMs excel in low-data image classification.
• Language knowledge enhances vision task performance.

Method

Image classification can be framed as text generation using VLMs, where the model outputs the class name given an image, fine-tuned with LoRA.

In practice

• Fine-tune Kimi K2 Thinking for complex reasoning.
• Integrate Tinker with OpenAI API-compatible tools.
• Use Qwen3-VL for image classification with limited data.

Topics

• Tinker Platform
• Kimi K2 Thinking
• OpenAI API Compatibility
• Vision-Language Models
• Image Classification

Code references

• thinking-machines-lab/tinker-cookbook

Best for: AI Architect, Computer Vision Engineer, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Tinker: Call for Community Projects — Tinker invites ML projects for publication, emphasizing rigorous evaluation and open-source contributions.
• Qwen3-VL — Qwen3-VL is a powerful multimodal model with broad visual and textual understanding capabilities.
• Last Week in AI #334 - Kimi K2.5 & Code, Genie 3, OpenClaw & Moltbook — Multimodal AI models and coding agents are rapidly advancing, alongside new interactive world-building and always-on AI assistants.
• Zamba2-VL Technical Report — Zamba2-VL combines Mamba2 and Transformers for competitive VLM performance with significantly lower inference latency, especially on-device.
• From DeepSeek V3 to V3.2: Architecture, Sparse Attention, and RL Updates — DeepSeek V3.2 combines sparse attention and advanced RL techniques for efficient, high-performing open-weight LLMs.
• How AI Vision Evolved | Merve Noyan — Vision models have evolved from CNNs to ViTs and now to scalable, open-source Vision-Language Models like LLaVA.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thinking Machines Lab - Connectionism.