Thinking Machines Releases Inkling-Small, a 276B Parameter Mixture-of-Experts Model

On July 30, 2026, Thinking Machines released Inkling-Small, an open-weights Mixture-of-Experts model with 276 billion total and 12 billion active parameters. Despite being a quarter of the size of the original 975-billion-parameter Inkling, Inkling-Small achieves comparable performance, scoring 31.6% on Humanity's Last Exam and over 80% on SWEBench-Verified. The model features native multimodal capabilities for audio and vision, and matches its predecessor on safety benchmarks like StrongREJECT.

Inkling-Small model release

• ▪Inkling-Small is available for fine-tuning on Tinker and for text, image, and audio chat on Tinker Playground as of July 30, 2026.
• ▪Thinking Machines released Inkling-Small, an open-weights artificial intelligence model, on July 30, 2026.

Mixture-of-Experts architecture efficiency

• ▪Inkling-Small is a Mixture-of-Experts transformer model featuring 276 billion total parameters and 12 billion active parameters.
• ▪Inkling-Small achieves comparable performance to the original Inkling model, which has 975 billion total parameters, at approximately one-quarter of its size.

Agentic coding task performance

• ▪Inkling-Small achieved a score exceeding 80% on the SWEBench-Verified benchmark using a bash-only harness.
• ▪Inkling-Small scored 31.6% on the Humanity's Last Exam benchmark, surpassing the original Inkling model's score of 29.7%.

Multimodal audio processing

• ▪Inkling-Small processes audio natively by representing it as dMel spectrograms within an encoder-free architecture.
• ▪Inkling-Small retains strong performance on speech understanding and long-form audio reasoning at a lower cost than the original Inkling model.

Vision reasoning capabilities

• ▪Inkling-Small processes images by dividing them into 40x40-pixel patches and transforming them using a four-layer hMLP.
• ▪Inkling-Small combines visual reasoning with Python operations like cropping, zooming, and programmatic image inspection to analyze documents and charts.

Safety benchmarking results

• ▪Inkling-Small is competitive in refusal and over-refusal rates on the FORTRESS benchmark, which measures safety across crime, violence, and dual-use risks.
• ▪Inkling-Small matches the performance of the original Inkling model on the StrongREJECT benchmark, which measures the refusal of harmful requests.

1 source