Inkling is a multimodal mixture-of-experts model with 975B total parameters and 41B active parameters per inference, a design that aims to keep individual request latency and compute cost manageable while retaining the representational breadth of a much larger model. The architecture draws on standard MoE routing, where only a fraction of the expert weights fire for any given token, making it well suited to long-context reasoning over mixed text, image, and audio inputs. This combination positions Inkling as a general-purpose model aimed at workloads that need cross-modal understanding without committing to a fully dense parameter budget on every call. Practically, the model fits teams building agents, retrieval pipelines, or analytic tools that must consume documents, screenshots, or audio alongside natural-language instructions. Its tool-use support allows it to call external functions as part of a reasoning chain, while the 256k context window gives it room to hold substantial transcripts, codebases, or multi-document retrieval results in a single prompt. The sparse active-parameter footprint also suggests a good balance between depth of reasoning and operational throughput, which is helpful for production deployments where cost-per-call and response time matter as much as raw answer quality.
From an integration standpoint, Inkling is surfaced through the Vercel AI Gateway under the namespace "thinkingmachines/inkling," with explicit capability tags for reasoning, tool use, vision, and image file input. That makes it straightforward to swap into existing gateway-based applications, since the model accepts both direct text prompts and structured image attachments within the same call. Because it inherits the gateway's standard streaming and provider-routing interfaces, teams already using AI Gateway infrastructure can adopt Inkling without bespoke client code. The model is best understood as a reasoning-oriented multimodal workhorse rather than a specialized chat or generation model. Its strengths align with tasks such as grounded question answering over mixed media, code and document analysis where diagrams or screenshots are part of the input, and agentic workflows that require the model to invoke tools and reflect on intermediate results. For practitioners evaluating Inkling, the practical fit is in applications where sparse-expert efficiency, long context, and the ability to reason across modalities outweigh the need for the absolute largest active-parameter budget on every request.
