AI-generated concept illustration: a modular sculpture and closed transport case symbolize hosted access and the pending weight release
There are two release dates to watch for Inkling 2 Small. The first has arrived: Thinking Machines Lab announced the model on November 5, with access through Tinker Playground and hosted fine-tuning and sampling. The second is still a plan. The lab intends to publish full weights under Apache 2.0 the following week, after external safety testing and red-teaming are complete.[1]
The launch announcement describes a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. It accepts images and supports variable reasoning effort. The company reports a score of 31 on the Artificial Analysis Intelligence Index; that figure is attributed to the lab’s announcement here, rather than an independently reproduced evaluation.[1]
Thinking Machines also identifies practical limitations: responses can be verbose, tool choices can be weak in codebases containing many skills, and conversational presentation needs improvement. Audio is not supported. The announcement says downloadable checkpoints will accompany the weight release.[1]
For developers, the editorial significance is the gap between evaluating a service and taking responsibility for a deployment. A hosted test can establish whether a model understands an internal task. It does not yet establish how the same workload will behave on a team’s own infrastructure. Downloadable weights, when delivered, would make that second investigation possible; they would not supply its answer.
The word “Small” also deserves a practical reading. A sparse model can use a fraction of its parameters for each token while still requiring a deployment plan for the full model. Teams considering local operation should distinguish active computation from memory, storage and serving requirements. This is an engineering consideration, not a measured hardware requirement for this release.
A useful evaluation would therefore start with an ordinary piece of work: a repository change that must pass existing checks, or a document task whose conclusions can be traced to the supplied material. Count the human corrections and unnecessary tool calls alongside the successful output. Variable reasoning effort is most useful when teams can see what extra computation buys them on those tasks.
The release also invites a closer look at tool selection. Adding more available skills can make an agent’s environment richer while making its choices harder to inspect. A trial should record which tools were offered, which were used and whether the task actually required them. These are proposed evaluation criteria, not results from an AIYXL test of Inkling 2 Small.
As of this review, the confirmed milestone is the hosted launch. The planned weight release remains conditional on testing, with no exact calendar date promised in the announcement. That distinction lets developers begin evaluation now without treating a future deployment option as an asset already in hand.
Sources:
