Inkling Small
Neural Network
Inkling Small is an open multimodal MoE model by Thinking Machines Lab with 12B active parameters out of 276B and API access.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 524 288 tokens
- Max answer length
- 262 144 tokens
- Reasoning effort
- none, minimal, low, medium, high, max
- Capabilities
- reasoning, function calling
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works Inkling Small?
Inkling Small is an open-weights multimodal model by Thinking Machines Lab with a mixture-of-experts architecture: out of 276 billion parameters, about 12 billion are active per request, reducing compute cost per token. It is a more compact and economical option in its lineup—ideal when you need reasonable processing speed without sacrificing multimodality. Via BotHub, the model is accessible from Russia without a VPN or foreign card, with pay-as-you-go billing in rubles: no subscriptions, and unused tokens do not expire. With over 250 neural networks in one window, you can compare responses and switch models in a few clicks, while a unified OpenAI-compatible API eliminates the need to rewrite integrations. Dialogues are encrypted via AES-GCM and not saved. Companies have access to contracts, invoices, EDI, an admin panel, and employee limits. Useful if you are building a multimodal assistant, analyzing text with images in a single request, preparing summaries of long materials, processing batch requests, or testing open models before choosing.Frequently asked questions about Inkling Small
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
inkling-small is a text model from Thinking Machines with reasoning before answering and tool calling. It accepts text, documents, and images, making it suitable for analyzing long materials, working with code, analytics, and scenarios requiring step-by-step output rather than a quick template response.
In a single response, the model outputs up to 262,144 tokens—this is a very long text: a large report, documentation, or an entire voluminous code file. The context window is 524,288 tokens, so source files and the response fit into one request without splitting.
Yes, reasoning before answering is a stated capability of the model: before providing a result, it breaks down the task step-by-step. This is noticeable in logic, mathematics, code analysis, and multi-step instructions where it is important to arrive at the answer sequentially rather than guessing.
Yes, the model supports function calling and structured JSON output. You describe the tools, it decides when to apply them, and returns the answer in the specified schema—convenient for agents, document parsing, and integrations where the result goes directly into code without manual parsing.
Images and documents—yes: send a screenshot, diagram, scan, or PDF, and the model will analyze them along with your question. Audio and video are not listed as input formats in our data. The model responds with text; generation of other formats is not supported.
We do not have data on the quality of performance for specific languages, so we will not make any promises in advance. The easiest way to check is in practice: send your task to BotHub, compare the response with other models in the same window, and choose the suitable one. Access from Russia without VPN, payment in rubles.