A Guth Labs publication ↗
Back to News

Thinking Machines Lab releases open weights for Inkling 2 Mini

A translucent green spiral wraps an ivory mechanism with one gold segment.
AI-generated editorial illustration

Thinking Machines Lab released Inkling 2 Mini on November 20, publishing its weights under Apache 2.0 and making it available for hosted use and customization through Tinker. The model accepts text and images, produces text and supports a context window of up to one million tokens.

Inkling 2 Mini has 26.4 billion parameters, with 3.4 billion active. It runs its middle layers three times per token, reusing the same weights and bringing the effective parameter activations to 5.4 billion per token. The model card lists a roughly 20 GB checkpoint and support for Apple silicon devices with at least 32 GB of memory and NVIDIA DGX Spark.

In the lab’s published results, Mini scored 48.2% on DeepSWE v1.1. It trailed Qwen3.8-27B on Terminal Bench v2.1, and the lab notes that evaluation setups vary by model. The release gives developers who cannot run the much larger Inkling 2 Small a way to test this architecture on supported hardware.

Sources and citations

Revision history

Published .