Thinking Machines Lab Ships Inkling: 975B-Param Open MoE, Tunable Reasoning

Thinking Machines Lab has released Inkling, a 975B-parameter (41B active) open-weights multimodal MoE model. Unlike typical omni models, Inkling skips separate encoders entirely: audio is converted into dMel spectrograms and images into 40x40 pixel patches via a four-layer hMLP, with both feeding into the same decoder alongside text.

Instead of RoPE, the model uses relative attention that encodes position directly in attention logits through a dedicated per-token, per-head projection. It stacks 66 layers and supports a 1M-token context window. The MoE layer routes over 256 experts (top-6) plus 2 always-active shared experts using sigmoid-based routing without auxiliary load-balancing losses.

The standout feature is a trained 'reasoning_effort' control: during RL, the team varied system prompts and per-token cost, teaching the model to self-regulate its token budget. This is now exposed as a parameter in transformers, letting Inkling match Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly a third of the tokens.

Benchmark highlights at effort=0.99 include 78.0% on FORTRESS Adversarial, the highest among compared open-weights models, 77.6% on SWEBench Verified, and 63.8% on Terminal Bench 2.1, trailing GLM 5.2 by 18.9 points. For engineers, the encoder-free design and controllable reasoning effort offer a practical lever for tuning cost versus accuracy in production deployments.