Baseten has made Thinking Machines Lab's Inkling Small model available through its hosted Model APIs and as a dedicated deployment. The release provides a managed route to a smaller member of the Inkling family after Baseten added the original 975-billion-parameter model in July.
Inkling Small is an open-weight mixture-of-experts model with 276 billion total parameters and 12 billion active parameters. It accepts text, image and audio inputs, returns text and supports a context window of up to one million tokens.
How developers can access the model
Developers can call Inkling Small through Baseten's OpenAI-compatible endpoint using the model identifier thinkingmachines/inkling-small and a Baseten API key. OpenAI compatibility can reduce integration work for applications already using the familiar chat-completions request pattern, although teams should still check parameter support and response behaviour before switching models.
Baseten is also offering dedicated deployments for organisations that need isolated capacity or more predictable operational settings. The announcement does not provide detailed service-level terms, regional availability or pricing, so those factors need to be confirmed for production planning.
The hosted launch is distinct from Baseten's earlier access to the full Inkling model. Inkling Small has roughly a quarter of the total parameter count and fewer active parameters per token. Thinking Machines Lab positions it as delivering similar performance to its larger sibling at a substantially smaller size, but workload-specific testing is needed to determine where that claim holds.
Why the smaller model may matter
A smaller sparse model can reduce the infrastructure needed to serve each request while retaining a broad set of capabilities. That can make it more practical for interactive applications, development environments and workloads where the full model's memory footprint is difficult to justify.
The one-million-token context window creates room for large document collections, long conversations and extensive code repositories. Context capacity is not the same as dependable recall or reasoning across the entire input, however. Teams should measure accuracy at different context lengths and compare retrieval-based approaches where latency and cost matter.
Native image and audio inputs broaden the potential uses beyond text-only chat. Developers could use the model to analyse visual material or spoken content alongside written instructions, but the announcement does not specify all supported media formats, size limits or processing charges.
What teams should evaluate
Before production use, teams should benchmark output quality, time to first token, sustained throughput and total cost against both the full Inkling model and smaller alternatives. A useful comparison should include representative multimodal inputs, tool calls, structured outputs and long-context cases rather than relying on headline parameter counts.
Open weights provide additional deployment and customisation options, but Baseten's hosted service is a managed access path. Organisations that require weight portability, fine-tuned checkpoints or self-hosting should confirm the applicable model licence and how artefacts can move between Thinking Machines Lab's tooling, Baseten and their own infrastructure.
Security and data handling also need review. Buyers should confirm retention settings, logging, regional processing, access controls and whether image or audio inputs receive different treatment from text. Dedicated infrastructure may change the isolation model, but it does not remove the need for application-level permissions and audit trails.
Baseten's release gives developers a quicker way to compare the two Inkling sizes without deploying either model themselves. Its value will depend on measured quality per dollar and how effectively the smaller model preserves the original's multimodal, reasoning and long-context behaviour under real traffic.