Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort
975B parameters. 41B active. Thinking Machines Lab just open-sourced Inkling—and deliberately positioned it as a customization base, not a capability leader.
Why it matters
Open-weights multimodal MoE with controllable thinking effort represents a shift in positioning: strength-through-adaptability over raw benchmark performance. Signals emerging bet on fine-tuning and inference control as differentiators in commoditizing model releases.
The key facts
9 to know975B-parameter Mixture-of-Experts architecture
41B active parameters
1M-token context window
Native multimodal: text, image, audio
Apache 2.0 open-weights release
Controllable thinking effort as positioning differentiator
Lab explicitly states model is not strongest available
July 15, 2026 release date
First model trained from scratch by Thinking Machines Lab
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Thinking Machines Lab released Inkling on July 15, 2026, its first model trained from scratch. The full weights ship under Apache 2.0. It is a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, a 1M-token context window, and native text, image, and audio input. The lab states…