Thinking Machines Lab introduced Inkling 2 Small on 5 November 2026 for use through its Tinker platform. Developers can run inference and fine-tune the model there now. Its full weights remain closed. The company intends to release them under Apache 2.0 after external safety testing, targeting the following week.
That distinction matters for teams choosing between a hosted evaluation and self-hosting. The Tinker release lets developers test the model today, but it does not provide weights to inspect or deploy independently. Thinking Machines Lab says fine-tuned checkpoints will become downloadable when the weights are released. Neither download is available as part of the current announcement.
What developers can use now
The model card describes a general-purpose model that takes text and images as input and produces text. It lists a context window of up to one million tokens. That published limit describes supported capacity, not the quality, speed or cost a particular application will see near the limit.
Tinker provides API access for inference and fine-tuning, while its playground offers an inference-only way to try the model. The current model-card licence is the Tinker Terms of Use. Teams evaluating it should check the applicable platform terms and test their own prompts, tools and data before treating the future Apache 2.0 plan as a present deployment option.
The model does not support audio input. That makes its multimodal scope narrower than a general claim that it can handle every media type would suggest.
What the parameter counts mean
Thinking Machines Lab describes Inkling 2 Small as a 48-layer, sparse mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. Its model card says each token is routed to eight of 384 experts. “Active” refers to the portion used for a token; it does not mean that the full model contains only 17 billion parameters.
That architecture gives developers a useful way to understand the announcement, but it does not establish a self-hosting hardware requirement or inference cost. Those questions need the actual released artifacts, deployment configuration and workload measurements. For now, Tinker is the stated access route.
What still needs testing
The company publishes coding, visual and other evaluation results, but benchmark tables are not a substitute for an application’s own tests. The announcement itself identifies weak spots: the model can choose tools poorly in codebases with many skills, be verbose, and organise chat responses less effectively than desired. The model card also warns of hallucinations and instruction-following failures.
A practical evaluation should therefore include the team’s real tool calls, long-context cases, image inputs and failure handling, with human review where an incorrect answer would matter. The open-weight release and checkpoint downloads should be treated as planned milestones until Thinking Machines Lab publishes the files and terms.
