Thinking Machines Lab introduced Inkling 2 Small on November 5, opening access through its Tinker platform while keeping the model's weights closed. Developers can try it and fine-tune it now, but downloading the weights remains a separate, conditional release. The company says it intends to publish them under Apache 2.0 next week, after external safety testing and red-teaming finish.
That timing matters if you're choosing a model for infrastructure you control. Hosted experimentation is available today. A deployment plan that depends on possessing the weights still has to wait. This report examines the announcement and documentation; we haven't run the model or verified its performance claims.
What developers can access now
Tinker offers sampling, meaning generating responses, and fine-tuning, which adapts the model through training. Its playground provides inference-only access, according to the model card. The launch announcement says fine-tuned checkpoints will become downloadable when the model's weights are released. Training a checkpoint on Tinker and taking it elsewhere are therefore separate steps in this launch.
Production access needs another check. Tinker's documentation labels serverless inference as beta and directs developers interested in production use to a waitlist. It asks for expected volume, latency requirements and the intended use case. Playground availability alone doesn't settle whether the service fits a production workload.
What “Small” describes
The model card lists 594 billion total parameters, with 17 billion active per token. Inkling 2 Small uses a mixture-of-experts architecture, routing each token to eight of 384 experts. That explains the gap between its total and active parameter counts. The smaller active count describes how much of the model participates in processing a token; it doesn't turn the full model into a 17-billion-parameter download.
The documented input and output are text and images in, text out. Thinking Machines says the model doesn't support audio. For a developer assessing it for a voice application, that is an immediate scope limit, regardless of its coding results.
Tool use still needs checking
Thinking Machines positions the model for extended coding and research tasks, but its own limitations section says it sometimes chooses tools poorly, especially in codebases with many skills. It also describes overly verbose responses and weaker organization in chat. Those admissions give developers specific behavior to test before handing an agent a large tool collection.
The next concrete milestone is the promised weight release, including the Apache 2.0 license and downloadable fine-tuned checkpoints. Until those artifacts arrive, developers can evaluate Tinker access and tool selection, while keeping any plan for running the model themselves contingent on the actual release.