Thinky released Inkling, a 975B-parameter mixture-of-experts transformer that the company positions as its first fully open-weights foundation model.

The model uses 41B active parameters and supports text, image, and audio inputs with up to 1M token context windows. Thinky trained Inkling from scratch on 45 trillion tokens across multiple modalities and released it under Apache 2.0 licensing.

Inkling debuts at 41 on Artificial Analysis's Intelligence Index, making it the highest-ranking US-based open-weights model. It outperforms Nemotron 3 Ultra (38), Gemma 4 31B (29), and other American open models, though it trails top Chinese releases like GLM-5.2 and Kimi K2.6 on some benchmarks.

Technical architecture sets it apart

The model incorporates several unusual design choices for a large-scale release. Thinky used relative positional encoding instead of the more common RoPE approach, and added short convolution layers around attention streams.

The mixture-of-experts setup includes two shared experts rather than the typical single shared expert. The company also implemented hybrid sliding-window attention with a 5:1 local-to-global layer ratio.

Inkling includes controllable reasoning effort levels, allowing users to adjust computational intensity based on task requirements. Early users highlighted its concise reasoning style and strong tool-calling capabilities.

Thinky simultaneously released Inkling-Small, a 276B-parameter variant with 12B active parameters that achieves competitive performance at lower computational cost.

The model launched with broad ecosystem support including vLLM, SGLang, Modal, Baseten, and Databricks. It entered Design Arena's Agentic Web App Arena at rank 9 with an Elo of 1257, placing it alongside Claude Opus 4.6 and Gemini 3.5 Flash.

Thinky began pretraining last winter, with a small team adding coding and reasoning capabilities from mid-January. The company emphasized this as a foundation release rather than their final frontier model.