Tinker becomes generally available, adding a sampling tool to test models during fine-tuning

Thinking Machines Lab/Tinker

Thinking Machines opened Tinker to everyone, dropping the waitlist, and added Kimi K2 Thinking as a model option. A new OpenAI-compatible interface lets users sample from models while training, and Qwen3-VL vision models now support image input.

• Supports large mixture-of-experts models and uses LoRA to share compute across multiple fine-tuning jobs, cutting infrastructure costs.
• New Qwen3-VL models handle 30B and 235B parameters for vision tasks including image classification.
• Inference sampling during training via OpenAI-compatible API reduces iteration time for researchers.