Inkling 2 Small has been added to the Workers AI catalog. The model adds multimodal input, a massive context window, and configurable reasoning effort, which directly affect how AI engineers, platform teams, and SREs design and operate their services.
What Changed
The new model, identified as @cf/thinkingmachines/inkling-2-small, is a second‑generation Inkling offering from Thinking Machines Lab. It features 594 B total parameters, 18 B active parameters, and a Mixture‑of‑Experts architecture. Key technical additions include:
- Vision support: accepts both text and image inputs in a single request.
- Reasoning control: default reasoning set to
higheffort; can be adjusted tomax,medium,low,minimal, or turned off withnone. - Function calling: enables agents to invoke external tools or APIs across multiple turns.
- Extended context: a full 1,048,576‑token window.
Why It Matters to Practitioners
For AI engineers, the larger context window reduces the need for chunking large documents, simplifying prompt engineering. Vision capability opens the door to image‑augmented workflows without a separate service. The configurable reasoning effort lets teams balance latency against depth of analysis. Platform and DevOps teams must now account for higher token consumption and the associated cost model, as the model is only available under the Workers Paid plan or with prepaid AI Gateway credits.
Architectural and Operational Implications
Deploying Inkling 2 Small requires a paid Workers subscription or prepaid AI Gateway credits, which introduces a billing dependency that must be reflected in budgeting and monitoring. The default rate limit is 20 requests per minute; it rises to 50 rpm when using prepaid credits. Engineers should design request throttling or queueing to stay within these limits, especially for high‑throughput agents.
Because the model accepts image data, services must handle binary payloads securely and ensure that any downstream storage complies with data‑handling policies. The expanded context window can increase memory pressure on Workers runtimes; testing for runtime memory usage is advisable before scaling.
Function calling across multiple turns means that state management becomes more critical. SREs should monitor session continuity and consider idempotency when agents invoke external APIs.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Evaluate whether Inkling 2 Small’s vision and long‑context features replace existing pipelines, potentially reducing architectural complexity. Update rate‑limit handling and budgeting to reflect the paid‑plan requirement. Test memory footprints and latency under realistic token loads, and verify that image handling complies with your security posture. Finally, monitor the reasoning‑effort setting to tune performance versus depth of analysis for your specific workloads.
