Skip to content

Introducing Inkling 2 Small

Thinking Machines Lab

The 594B-parameter model is available on Tinker for inference and fine-tuning, with weights scheduled for Apache 2.0 release next week.

Key points

  • Activates 17B parameters per token and includes a 292M-parameter vision encoder for image inputs.
  • Pretrained across four phases on over 12.5 trillion tokens in native NVFP4 on NVIDIA GB300 NVL72 hardware.
  • Trained via a joint reinforcement learning run over 1,075 steps and 17 million rollouts across coding, STEM, and agentic workflows.
  • Scored 31 on the Artificial Analysis Intelligence Index and 1527 on AA-Briefcase for knowledge work tasks.

Why it matters

Developers using Tinker get immediate access to a multimodal MoE model for agentic coding and research, with open weights arriving next week.

Sources