Thinking Machines ships Inkling, a 975B open-weights MoE built for fine-tuning

Thinking Machines has released Inkling, a Mixture-of-Experts model trained from scratch with full weights published for anyone to adapt. The model carries 975B total parameters with 41B active per token, handles context windows up to 1M tokens, and was pretrained on 45 trillion tokens spanning text, images, audio, and video. A lighter preview, Inkling-Small, runs 12B active parameters for lower cost and latency. The company is candid that Inkling is not the strongest model available, open or closed — the pitch is that its native multimodality, controllable thinking effort, and immediate availability for fine-tuning make it a strong base to build on rather than a benchmark leader.

Technically, the standout choices are efficiency and an encoder-free multimodal design. Audio enters as discrete dMel spectrograms and images as 40×40 pixel patches, both fed through a lightweight embedding layer and processed jointly with text rather than through separate vision or audio encoders. Inkling’s tunable reasoning budget lets developers trade tokens for accuracy; the company claims it matches Nemotron 3 Ultra on Terminal Bench while spending a third of the tokens. It was also trained across randomized tool schemas to reduce brittleness inside different agent harnesses, and ranks competitively among open-weight models on agentic coding and web-dev leaderboards like Design Arena.

The release is tied directly to Thinking Machines’ Tinker platform, where Inkling is available for fine-tuning today alongside a new in-console Playground for interacting with the model. The company frames it as the reasoning backbone for its real-time, voice-and-vision interaction models, and emphasizes ‘epistemics’ — calibration, instruction-following, and resistance to censorship — as design goals, pointing to forecasting as a use case where fine-tuned models have been outpacing frontier LLMs. Inkling is positioned as the first entry in an ongoing model family.