Inkling 2 Small
A large language model with a mixture-of-experts (MoE) architecture released by Thinking Machines Lab. It has 594B total parameters and 17B activated parameters, is served through the Tinker platform, and is planned for release under the Apache 2.0 license with open weights.
Core Facts
The official release information does not provide specific benchmark data, context length, or quantitative comparisons with competing products.
For coding agent tools such as Claude Code, Codex, and Cursor, you need to first install the latest version of Vercel CLI and run the 'vercel ai-gateway setup' configuration command.
Inkling 2 Small has a total parameter count of 594B, activating 17B parameters per inference.
Timeline (last 90 days)
Thinking Machines' Inkling 2 Small model is officially live on the Ollama platform.
The contrast between the "Small" naming and the 594B scale may suggest that Thinking Machines also has larger-scale versions planned, or that "Small" refers to its positioning relative to its other internal models.
For a 594B-parameter model, the hardware bar for local running is not low, and calling it via the cloud is the more realistic choice.
The command to access Inkling 2 Small through Ollama is `ollama run inkling-2-small:cloud`.
Inkling 2 Small's official example use cases include OCR, grounding, infographic interpretation, visual reasoning, tool calling via MCP, financial spreadsheet reasoning, and earnings report and document processing.
Inkling 2 Small claims strong performance in visual interpretation and creation, agentic knowledge work, and financial and quantitative reasoning.
Inkling 2 Small offers a Zero Data Retention commitment on Ollama Cloud.
Inkling 2 Small is a 594B-parameter mixture-of-experts (MoE) model equipped with a 1M context window.
Thinking Machines disclosed the training method for Inkling 2 Small: forward computation used 4-bit precision throughout pretraining, pretraining took only 68 hours, and the weights have been open-sourced.
Thinking Machines open-sourced the weights of Inkling 2 Small under the Apache 2.0 license, and developers can download the weights and deploy them themselves.
39 more timeline events
All Facts (20)
The official release information does not provide specific benchmark data, context length, or quantitative comparisons with competing products.
75%VerifiedFor coding agent tools such as Claude Code, Codex, and Cursor, you need to first install the latest version of Vercel CLI and run the 'vercel ai-gateway setup' configuration command.
65%Partially VerifiedInkling 2 Small has a total parameter count of 594B, activating 17B parameters per inference.
65%UnverifiedInkling 2 Small offers a Zero Data Retention commitment on Ollama Cloud.
50%UnverifiedInkling 2 Small claims strong performance in visual interpretation and creation, agentic knowledge work, and financial and quantitative reasoning.
50%UnverifiedInkling 2 Small is a 594B-parameter mixture-of-experts (MoE) model equipped with a 1M context window.
50%UnverifiedInkling 2 Small's official example use cases include OCR, grounding, infographic interpretation, visual reasoning, tool calling via MCP, financial spreadsheet reasoning, and earnings report and document processing.
50%UnverifiedFor a 594B-parameter model, the hardware bar for local running is not low, and calling it via the cloud is the more realistic choice.
50%UnverifiedThe command to access Inkling 2 Small through Ollama is `ollama run inkling-2-small:cloud`.
50%UnverifiedThe native 4-bit version of Inkling 2 Small requires at least 500GB of aggregate VRAM.
50%UnverifiedGemma 4 12B abandoned the vision encoder, Kimi K3 dropped encoder pretraining, while Inkling 2 Small retained the encoder.
50%UnverifiedOversized MoE models that support multimodal image input and are licensed under Apache 2.0 remain scarce in the open-source community.
50%UnverifiedInkling 2 Small is a mixture-of-experts (MoE) model with 594B total parameters and 17B activated parameters, and supports image input.
50%UnverifiedEven with only 17B activated parameters, the full weights of 594B total parameters still require considerable VRAM or memory to hold, which is difficult to fit on an ordinary single GPU; actual operation most likely requires multi-GPU parallelism or a high-memory configuration.
50%UnverifiedIn visual development scenarios, Inkling 2 Small can use screenshots to check the code output it has written and use that to guide the next round of revisions.
50%UnverifiedWhen calling Inkling 2 Small, use the model name thinkingmachines/inkling-2-small.
50%UnverifiedInkling 2 Small is positioned to cover scenarios such as coding agents, visual understanding, and professional knowledge work.
50%UnverifiedInkling 2 Small is a multimodal reasoning model that supports text and image input and text output
50%UnverifiedInkling 2 Small provides inference and fine-tuning services through the Tinker platform.
50%UnverifiedThe official Product Hunt page does not disclose the specific implementation of "can see and optimize" (such as whether a visual encoder or multimodal input is built in).
50%