Inkling 2 Small: a “small” model with 594B parameters

Only the active portion is “small”

What stands out about Inkling 2 Small is not its size but its use of the word small. In its announcement and model card, Thinking Machines AI describes Inkling 2 Small as an open-weight multimodal model with 594 billion parameters, while only 17 billion are active for each token.

It uses a sparse Mixture-of-Experts architecture: every token is routed to 8 of 384 experts. The architecture is a 48-layer decoder-only Transformer with hybrid local and global attention. The premise is straightforward: retain a large total capacity without paying to compute every parameter for every token.

Still, 17 billion active parameters do not mean the model can be deployed like a conventional compact 17B model. All 594 billion weights must still be stored or distributed across devices. Compute savings per token therefore coexist with substantial requirements for memory, expert routing, and inter-server communication.

The model accepts text and images, with a claimed context window of roughly 1 million tokens. Reasoning effort is configurable; the published evaluations used a 0.99 setting labeled xhigh. That detail matters: quality cannot be separated from the compute budget and the length of the generated answer.

The model card reports 82.2% on SWEBench Verified, 85.1% on Terminal-Bench 2.1, and 81.6% on MMMU Pro. These are developer-reported results, and Inkling 2 Small does not lead every comparison. As of November 6, 2026, independent validation remains an especially important part of the picture.

Sparsity does not guarantee a cheap final answer

The practical takeaway is that Inkling 2 Small may reduce compute for an individual token without necessarily reducing the cost of a complete task. According to the published data, the model generated about 150,000 output tokens per task on average—roughly four times as many as its predecessor, Inkling Small.

That is where the arithmetic becomes less impressive. If reasoning expands, the benefit of 17 billion active parameters is partly consumed by long trajectories, context processing, and infrastructure for 594 billion weights. I would look first at the total cost of a completed task and the latency until a useful result, not the active-parameter count alone.

For researchers and teams with distributed infrastructure, the model is meaningful: open weights, image input, long context, and strong claimed coding results create room for customization. For local or budget inference, however, the word Small is misleading. This is not a small model; it is a large model that has learned to switch itself on in parts.

We previously reviewed Pony Alpha on OpenRouter—likely GLM-5 with a 200K context window and capabilities for safe testing. That case usefully complements an assessment of Inkling 2 Small when comparing new models by context length and practical use cases.