Thinking Machines releases Inkling 2 Small on Tinker Thinking Machines releases Inkling 2 Small on Tinker

The size behind the name

The model card describes a sparse mixture-of-experts transformer with 48 layers, routing each token through eight of 384 experts. Active parameters describe computation for a token; storage must accommodate the full model. Text and images share a representation. Developers therefore face separate decisions about input support, context length and eventual infrastructure. A smaller active count alone does not answer all three.

Customization comes before export

The launch gives Tinker users a path to test and adapt the model without first operating their own inference infrastructure. The announcement says downloadable checkpoints will arrive alongside the full weights. That makes training access and model portability separate milestones. Thinking Machines intends an Apache 2.0 weight release. For teams that already use Tinker, this preserves the service as the place where an initial evaluation and a customization job can happen.

A change from the first Inkling

The July 15 Inkling announcement presented a different starting point: full weights were available at launch. That model had 975 billion total parameters and 41 billion active parameters, with native text, image and audio reasoning. Its stated context window already reached one million tokens. Hosted customization therefore predates this release; it is not a new capability invented for the second generation. The older announcement also positioned the family as a base for adaptation across workflows rather than the strongest model on every task. The new release makes the access sequence more important than the family label alone. The original launch demonstrated a model writing a fine-tuning job, running it through Tinker and evaluating the adapted behavior. It also described training across different agent harnesses and randomized tool schemas. Those choices explain why customization belongs in the product story alongside benchmark gains: the intended workflow includes adapting behavior, not simply replacing one chat endpoint.

Why the release is staged

Thinking Machines' July 31 policy explains why it separates hosted access from public weights. Once distributed, weights cannot be recalled from everyone who downloaded them. Its framework weighs both model capabilities and the surrounding defensive ecosystem. Monitored access and hosted fine-tuning are intermediate steps in that approach. The policy also treats removable refusal behavior differently from durable limits on harmful capabilities. That distinction changes what an open release promises: access to inspect and adapt a model comes with fewer controls over subsequent use. The policy makes testing a release condition, rather than treating a calendar target as sufficient evidence.