From the source
Baseten has added Thinking Machines' Inkling-Small to its Model APIs and Dedicated Inference, giving customers a smaller, cheaper way to run the same multimodal capabilities as Inkling.
- The model keeps Inkling's text, image and audio inputs and its 1M token context window, so Baseten describes it as an ideal drop-in replacement.
- At roughly one-quarter the size of Inkling, it uses a sparse Mixture-of-Experts architecture that lowers VRAM requirements, which reduces deployment costs, allows use on fewer GPUs and leaves headroom for faster inference.
- Baseten says Inkling-Small delivers comparable performance to Inkling across reasoning, coding, vision, audio and safety, and slightly outperforms it on several benchmarks, though individual scores fluctuate.
- Like Inkling, it is open-weight, and Baseten runs it on
NVIDIA accelerated infrastructure.
- Baseten says its Delivery Network gives 2–3× faster cold starts, and that multi-cloud capacity management keeps applications online when specific cloud resources are constrained.





