Thinking Machines: Introducing Inkling

Thinking Machines Lab announced Inkling — its open-weights model — via @thinkymachines.

#What it is

• 975B total parameters, 41B active — a Mixture-of-Experts transformer.
• Natively multimodal: text, images, audio, and video, over a 1M-token context window.
• Pretrained on 45 trillion tokens.

#What it's built for

• Agentic coding with tool use.
• Controllable reasoning effort — dial the compute/quality trade-off per task.
• Strong multimodal performance and calibrated uncertainty (it expresses how sure it is).
• Positioned as "a practical multimodal foundation model for customization across domains" rather than a single-benchmark optimizer.

#Availability

• Fine-tunable on Thinking Machines' Tinker platform.
• Inference via deployment partners: TogetherAI, Fireworks, Modal, and others.
• A smaller Inkling-Small (12B active) is previewed for latency-sensitive applications.

#Why it matters

An open-weights, natively-multimodal MoE at this scale — with per-task reasoning-effort control and calibrated uncertainty — is a serious entry in the open frontier, and the Tinker + multi-provider story makes it something teams can actually customize and deploy rather than just benchmark.

Sources:

• Introducing Inkling — Thinking Machines Lab
• @thinkymachines announcement