Thinking machines releases inkling small model

Thinking Machines has expanded its model family with the release of Inkling Small, a lighter-weight multimodal model that follows just two weeks after the launch of the flagship Inkling. Distributed under an Apache 2.0 license, the new model is designed to offer a more efficient footprint for developers while retaining much of the reasoning, coding, and multimodal capability of its larger sibling.

Architecture and Parameter Efficiency

Inkling Small is built as a sparse Mixture-of-Experts (MoE) decoder-only transformer featuring 42 layers. The architecture routes each token to 6 of 256 specialized experts, alongside 2 shared experts that remain active on every token. This design results in 276 billion total parameters with 12 billion active parameters per token, compared to the 975 billion total and 41 billion active parameters found in the original Inkling model.

The model natively accepts multiple input modalities, processing text via UTF-8 encoding, pixel-based images through a hierarchical patch encoder, and audio via discrete token encoding sampled at 16kHz. All modalities are projected into a shared hidden space and processed jointly by the decoder. While the flagship model supports a context window of up to 1 million tokens, Inkling Small operates with a 256K token context window.

Benchmark Performance

Despite having less than a third of the parameters of its predecessor, Inkling Small achieves competitive performance across a variety of evaluations. On the Artificial Analysis Intelligence Index, the model scores 40, coming within a single point of the flagship Inkling score of 41.

Furthermore, Inkling Small matches or exceeds the larger model on several key coding and reasoning benchmarks. It records an 80.2% score on SWE-bench Verified, compared to 77.6% for the flagship, and achieves 64.7% on Terminal Bench 2.1. It also scores higher on Humanity's Last Exam and GPQA Diamond evaluations. However, the model trails its predecessor on certain agentic and factual knowledge tasks, such as τ³-Banking.