Skip to main content
Back to timeline
CloudflareSource publication:

Thinking Machines' Inkling 2 Small lands on Workers AI: 594B total parameters, 18B active, and a 1,048,576-token context

Synopsis

Inkling 2 Small, the second generation of Thinking Machines Lab's Inkling models, is now available on Cloudflare Workers AI as @cf/thinkingmachines/inkling-2-small, the first frontier open model from a US lab on that platform. It is a multimodal Mixture-of-Experts model with 594B total parameters and 18B active parameters, accepting text and image inputs, with reasoning on by default, function calling across conversation turns, and a 1,048,576-token context window. Thinking Machines says improvements to pre-training and infrastructure let it train this generation much faster than previous models.

Source-provided article image: Changelog
Cloudflare

Interpretation

Workers AI now serves a frontier open model directly: Inkling 2 Small, the second-generation Inkling model from Thinking Machines Lab, offered under the name @cf/thinkingmachines/inkling-2-small. Before this, Workers AI had no frontier open model from a US lab; Inkling 2 Small is the first. The release states this position as the central fact of the launch and pairs it with the capability gain over the previous Inkling generation.

The model runs as a Mixture-of-Experts: it holds 594B total parameters but activates 18B of them, so inference uses only a fraction of the network. Mixture-of-Experts means the model is built from many expert subnetworks and routes each input through only some of them, and this generation also accepts images alongside text. Thinking Machines reports that pre-training and infrastructure improvements made this generation train much faster than previous models.

Callers can tune reasoning effort per task: it is on by default at high effort, and can be set to max, medium, low, minimal, or none to turn it off. Reasoning being on by default means the model reasons before answering unless the caller configures otherwise. The context window is 1,048,576 tokens, and function calling lets agents invoke tools and APIs across multiple conversation turns.

Perspective

The release addresses developers who will call the model on Workers AI, covering agentic coding, image understanding, function calling and a 1,048,576-token context; use requires the Workers Paid plan or prepaid AI Gateway credits.

The release gives no benchmark scores or specific comparison figures against the previous generation, and its capability descriptions are qualitative; the default rate limit is 20 requests per minute, or 50 requests per minute with prepaid AI Gateway credits.