Thinking machines unveils inkling 2 small, a sparse moe model with 594b parameters

Thinking Machines Lab has introduced Inkling 2 Small, the second generation model in its Inkling series. This multimodal Mixture-of-Experts (MoE) model features a total size of 594 billion parameters with 17 billion active parameters. It is designed for tasks such as agentic coding, vision, and knowledge work, accepting both text and image inputs, though it currently lacks audio support. On the Artificial Analysis Intelligence Index, Inkling 2 Small achieved a score of 31, significantly surpassing its predecessor, Inkling Small, which scored 13. The lab confirmed that the model's weights are currently closed, but an Apache 2.0 release is anticipated next week following external safety evaluations. Technically, the model incorporates a 48-layer decoder-only transformer, where each token routes to 8 out of 384 experts. The vision component utilizes a 292M-parameter Perception Encoder. Development involved a complex pretraining regimen across four phases, including a joint RL run that spanned 9 days on 2,048 GPUs, exceeding $4 imes 10^{23}$ FLOPs. Furthermore, the lab improved alignment by adding a reward and on-policy self-distillation, reducing its internal misbehavior score from 66.8 (on Inkling) to 14.9. The model is currently accessible on Tinker for fine-tuning and inference, with Artificial Analysis listing its pricing at $0.30 per 1M input tokens and $1.00 per 1M output tokens.