Tinker vs. Inkling

Tinker is a groundbreaking training API designed specifically for researchers and developers, granting them extensive control over model fine-tuning while alleviating the intricacies associated with infrastructure management. It provides fundamental building blocks that enable users to construct custom training loops, implement various supervision methods, and develop reinforcement learning workflows. At present, Tinker supports LoRA fine-tuning on open-weight models from the LLama and Qwen families, catering to a spectrum of model sizes that range from compact versions to large mixture-of-experts setups. Users have the flexibility to craft Python scripts for data handling, loss function management, and algorithmic execution, while Tinker efficiently manages scheduling, resource allocation, distributed training, and failure recovery independently. The platform empowers users to download model weights at different checkpoints, freeing them from the responsibility of overseeing the computational environment. Offered as a managed service, Tinker runs training jobs on Thinking Machines’ proprietary GPU infrastructure, relieving users of the burdens associated with cluster orchestration and allowing them to concentrate on refining and enhancing their models. This harmonious combination of features positions Tinker as an indispensable resource for propelling advancements in machine learning research and development, ultimately fostering greater innovation within the field.

Inkling is an open-weights multimodal AI model from Thinking Machines built to support customization, agentic workflows, coding, reasoning, vision, audio, and enterprise AI use cases. The model is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active parameters, 256 routed experts per MoE layer, and six routed experts active per token. It supports context windows up to 1 million tokens and was pretrained on 45 trillion tokens across text, images, audio, and video. Inkling is designed as a broad foundation model rather than a narrowly optimized benchmark model, giving it balanced capabilities across reasoning, coding, factuality, instruction following, vision, audio, tool use, and safety. Its controllable thinking effort lets developers adjust how much computation and generated reasoning the model uses, helping teams balance quality, latency, and cost for different production needs. The model can run agentic coding tasks, use tools, create web apps, generate polished multi-page artifacts, reason over long contexts, and work through iterative refinement loops. For multimodal tasks, Inkling can process images, answer questions about visual content, transcribe and reason over audio, follow spoken instructions, and combine visual reasoning with code-based tools such as Python. Thinking Machines trained Inkling for calibration, instruction following, factual reliability, refusal behavior, and safety across multiple modalities, including evaluations for dangerous capabilities and human-AI threat vectors. Inkling is available on Tinker for fine-tuning, with 64K and 256K context options, an Inkling Playground for testing, cookbook recipes, and support for multimodal post-training workflows. Its full weights are available on Hugging Face, and deployment support is available through APIs and infrastructure partners such as TogetherAI, Fireworks, Modal, Databricks, Baseten, SGLang, vLLM, llama.cpp, and transformers.