Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, making it available through Tinker for inference and additional training. The model has 594 billion total parameters and 17 billion active parameters. The company intends to release weights under Apache 2.0 after external safety testing. That is a future commitment: the current launch should not be described as weights already available for unrestricted download.
What developers can use
The model card specifies text and image inputs, text output, a context window of up to one million tokens, and distribution under Tinker's terms. This version does not support audio. Its sparse MoE architecture has 48 layers and routes each token to eight of 384 experts. These specifications describe the design; they do not establish application-level cost, latency, or accuracy.
The Inkling guide integrates the model into existing Tinker Cookbook workflows using thinkingmachines/Inkling-2-Small. Developers can control reasoning effort, and the documentation recommends comparing effort and learning-rate settings. That matters for teams seeking adaptation to a particular task rather than only a general conversational assistant.
Separate tml-renderers documentation explains how messages, tools, and tool results become model inputs. Training and generation need compatible representations. When the model proposes a tool call, the application inspects it, executes the function, and returns its result. A model output is not, by itself, authorization to act inside another system.
Editorial assessment: evaluate the complete loop
A useful comparison starts with a small, complete task. For example, an internal documentation assistant could receive a diagram image and propose a revision to its written explanation. An evaluation should check whether relationships were interpreted correctly, whether the answer relies on the actual image, and whether uncertainty is stated clearly. An elegant explanation alone does not establish accurate visual understanding.
The next level adds a tool. If the example application retrieves a document, reviewers should be able to identify the document and its version. If it saves a revision, a separate check should establish the resulting state. These are proposed editorial evaluation criteria, not a claim that SIExpo conducted its own comparative Inkling experiment.
Large context capacity requires its own assessment. Accepting many tokens does not establish that an agent consistently respects earlier requirements after successive changes. One useful test would include conflicting document versions and a question requiring the current version. The criterion should be set beforehand so a persuasive explanation cannot substitute for the correct outcome.
For additional training, adaptation examples should be separated from evaluation cases. Reusing the same cases in both can make memorization look like general improvement. Cost per completed attempt, repeated work, and human corrections provide a more useful picture than the price of a single short response. We report no independently measured performance or cost figure here.
An organization could keep the same small evaluation set while varying effort, recording whether extra reasoning actually changes the final result. A longer answer may be helpful, but length alone should not count as success. For a tool-using assistant, a completed, correctly checked change is a different outcome from an unfinished attempt that merely describes what should happen next.
Availability and limitations
The immediate practical development is a usable route to inference and customization through Tinker. Planned weights should be tracked as a separate event requiring actual files and an applicable published license. Until then, the decision concerns a service under its present conditions.
The official card warns about unsupported claims and uneven performance across languages and domains. A Bulgarian application therefore needs evaluation on its Bulgarian workflow. Our assessment is that the model merits targeted testing for multimodal and tool-based work; it does not justify automatically replacing an already validated system. We reviewed the release, card, and integration documentation, without installing the model or running a paid training job.
