Thinking Machines' Inkling-Small Squeezes a 276B Multimodal Model onto One GPU

Thinking Machines Lab has released Inkling-Small weights, an open-weight multimodal reasoning model with 276 billion total parameters and 12 billion active per token. It scores 40 on the Artificial Analysis Intelligence Index, one point behind the larger Inkling model despite using less than a third of its total and active parameters.

The model accepts text, images, and audio, supports variable reasoning effort, and offers a context window of up to one million tokens. Thinking Machines released the weights under Apache 2.0, allowing commercial use, modification, and redistribution subject to the license terms.

Smaller routing, similar aggregate scores

The original Inkling release uses a 975B-parameter mixture-of-experts architecture with 41B parameters active for each token. Inkling-Small reduces those figures to 276B total and 12B active, cutting the active parameter count by about 71%.

Mixture-of-experts models divide their feed-forward layers into specialized subnetworks called experts. Inkling-Small routes each token through six of 256 selectable experts and two shared experts, so the full 276B parameters do not run for every token. This design lowers inference compute while preserving a large pool of learned capacity.

The decoder-only transformer has 42 layers and alternates local attention with global attention to control the cost of processing long contexts. A hierarchical patch encoder converts images into model representations, while an audio tokenizer converts sound into discrete tokens. Both modalities enter the same hidden space as text, allowing the decoder to process them jointly. Training ran on NVIDIA GB300 NVL72 systems.

Coding and tool use lead the results

Thinking Machines reports that Inkling-Small matches or exceeds the larger model on several coding, agentic, and reasoning evaluations: