inkling

Inkling

Inkling is Thinking Machines Lab's open-weight model family built for Tinker: general-purpose models that code, reason, call tools, and accept image and audio input.

Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active. Inkling-Small is an efficient sibling at 276B total and 12B active, reaching comparable performance to Inkling at a quarter of its size. Both offer native reasoning over audio and images, variable thinking effort, a context window of up to 1M tokens, and well-rounded performance across a range of benchmarks.

Inkling-Small's efficiency makes it a reasonable default for most tasks, and a natural fit for workloads where cost and latency matter, such as coding, using LLMs to grade, or generating synthetic data for other models. The two share a renderer, tokenizer, and effort interface, so moving between them is a one-line change to the model name, and benchmarking both on your own task is cheap.

Both models are offered as post-trained versions on Tinker, not base models: they arrive instruction-tuned and effort-conditioned, so treat them as starting points for further post-training — SFT, RL, distillation — rather than for continued pretraining.

Working with Inkling differs from other Tinker models in three ways, each covered below: every render needs an explicit thinking-effort value, rendering and tokenization go through tml-renderers rather than a Hugging Face chat template, and the learning rate is yours to calibrate.

Setup

Inkling renders through the standalone tml-renderers package, included in the default installation together with the required torch>=2.10.

Pass thinkingmachines/Inkling or thinkingmachines/Inkling-Small anywhere the cookbook takes a model name. The tokenizer and renderer (tml_v0) are selected automatically for any thinkingmachines/Inkling* model, including the :peft: long-context variants: