Thinking Machines: Inkling Small

Thinking Machines: Inkling Small is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab. It has 12B active parameters out of 276B total, and it is positioned as the smaller, more efficient member of the Inkling family. The provider release date is 2026-07-30. Because the weights are open, teams that need to inspect, self-host or fine-tune a model have that option, while the hosted endpoint on api14 is available for those who would rather not run the infrastructure themselves.

The model accepts text, image and audio inputs and returns text. Its context window is 524288 tokens, and the maximum output is 262144 tokens, so it can take in very large documents, long recordings or image sets alongside a prompt and still produce a long response in a single call. The mixture-of-experts design means only a fraction of the total parameters are active per token, which is the main reason a model of this size can be served at reasonable latency and throughput. Reasoning behaviour is exposed through the API, so you can decide how much internal deliberation you want before the final answer.

Those capabilities point to a fairly specific set of uses. Long-context work is the obvious fit: summarising or querying large codebases, contract sets, research corpora or support ticket histories that would not fit in a smaller window. Multimodal input makes it useful for tasks that mix formats, such as extracting structured data from scanned documents with embedded images, transcribing and then reasoning over audio, or answering questions about a screenshot plus a written brief. Tool calling and reasoning support suit agent-style workflows where the model needs to decide which function to invoke and then explain its result.

You call it through api14's OpenAI-compatible API using the model id "thinkingmachines/inkling-small". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, seed, stop, temperature, tool_choice, tools, top_k and top_p. If you already have code written against the OpenAI chat completions shape, switching to this model is mostly a matter of changing the model string and, if you want reasoning output, setting include_reasoning or reasoning_effort.

It is a text-output model, so anything requiring generated images, audio or video is out of scope here. The 262144 token output cap is generous but still finite, which matters for very long generation jobs. If your workload is small, latency-sensitive and purely textual, a lighter model may be cheaper to run, and if you need output modalities beyond text, you will need a different model.