Thinky releases Inkling, a 975B-parameter Apache 2.0-licensed multimodal MoE model with 1M-token context

Thinky, a research lab that had previously disclosed limited details about its models, publicly released Inkling, its first fully open-weights foundation model family. The flagship model, Inkling, is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active parameters per token. It supports native multimodality across text, images, and audio, and offers a context window of up to 1 million tokens.

The company also introduced Inkling-Small, a lighter-weight variant with 276 billion total parameters and 12 billion active parameters, designed for lower cost and latency while maintaining competitive performance. Both models are released under the Apache 2.0 license, with open-weights checkpoints available immediately.

Training details shared by the company and corroborated by community analysis indicate Inkling was pretrained on approximately 45 trillion tokens spanning text, images, audio, and video. The release emphasizes controllable reasoning effort and numerical effort levels, positioning the models as customizable base models rather than benchmark-maximized systems.

Inkling is available on Thinky’s Tinker platform and Playground, with additional support from ecosystem partners including vLLM, SGLang, Modal, Baseten, Databricks, and Hugging Face. Community and third-party commentary highlighted Inkling as the leading U.S.-based open-weight model at launch, though still trailing top Chinese open-weight and best closed models on some benchmarks.

Independent evaluators noted Inkling’s strong performance on agentic tasks and token efficiency. Artificial Analysis reported the model debuted at an Intelligence Index score of 41, ahead of Nemotron 3 Ultra (38) and Gemma 4 31B (29), and described its average output length as 25,000 tokens per task, framing it as relatively token-efficient compared to peers.

Design Arena’s Agentic Web App Arena ranked Inkling at #9 overall with an Elo of 1,257, placing it in the same band as Claude Opus 4.6 and Gemini 3.5 Flash, and noted it as the highest-ranking U.S.-based open-weight model for agentic workloads.