Inkling-Small
Compact variant of Inkling for cost-conscious deployments, retaining the multimodal MoE architecture
Inkling-Small
Thinking Machines Lab • July 2026
Training Data
45 trillion tokens (text, image, audio, video)
Inkling-Small
July 2026
Parameters
276B total, 12B active (MoE)
Training Method
Muon/Adam pretraining + large-scale asynchronous RL
Context Window
1,000,000 tokens
Knowledge Cutoff
2026
Key Features
Multimodal (encoder-free) • Cost-Efficient • Controllable Thinking Effort • Open Weights
Capabilities
Efficiency: Outstanding
Multimodal: Very Good
Long Context: Excellent
What's New in This Version
Delivers Inkling's multimodal and agentic capabilities at a fraction of the compute for budget-sensitive deployments
Compact variant of Inkling for cost-conscious deployments, retaining the multimodal MoE architecture
What's New in This Version
Delivers Inkling's multimodal and agentic capabilities at a fraction of the compute for budget-sensitive deployments
Technical Specifications
Key Features
Capabilities
Other Thinking Machines Lab Models
Explore more models from Thinking Machines Lab
