Thinking Machines releases Inkling-Small with near-flagship performance

Thinking Machines Lab has released Inkling-Small, an open-weights AI model designed to deliver much of the performance of its larger Inkling model with substantially lower computing demands. The company says the model performs similarly to Inkling across reasoning, coding and multimodal tasks while using fewer active parameters for each generated token.

Inkling-Small has 276 billion total parameters, but activates 12 billion of them at a time. Its predecessor, Inkling, has 975 billion total parameters and activates 41 billion. Both models use a Mixture-of-Experts design, which directs each task to a limited set of specialised parts of the model instead of running all parameters for every token.

The release is available under the permissive Apache 2.0 licence. That allows organisations to use, modify and redistribute the weights, including in commercial products, subject to the licence terms. Thinking Machines Lab is also offering the model through its Tinker fine-tuning service and Tinker Playground chat interface.

Strong results, with important trade-offs

According to Thinking Machines Lab’s official announcement, Inkling-Small exceeds the larger model on several reasoning and agentic coding tests. The company reports a score of 31.6 percent on the text-only Humanity’s Last Exam benchmark, compared with 29.7 percent for Inkling. It also reports more than 80 percent on SWE-bench Pro, a benchmark for resolving software engineering tasks.

Carl Franzen reports for VentureBeat that Artificial Analysis gives Inkling-Small a score of 40 on its Intelligence Index, one point behind Inkling. The publication also cites company figures showing higher scores for Inkling-Small on Terminal Bench 2.1 and SWE-bench Verified.

Those results do not mean the smaller model wins across every category. Thinking Machines says Inkling retains an advantage in factual knowledge coverage and factuality. VentureBeat notes that Inkling-Small performs notably worse on the τ³-Banking benchmark, which tests complex tool use in a banking setting. Organisations using the model for factual or high-stakes work would therefore still need retrieval systems, source checks and human review.

Inkling-Small accepts text, image and audio inputs and produces text. It supports context windows of up to one million tokens. The company says it can reason over charts, documents, images and spoken material, and can use Python tools for visual inspection tasks such as zooming or cropping images.

“Small” remains relative

Despite its name, Inkling-Small is not intended for typical personal computers. VentureBeat reports that the standard BF16 version requires about 600 GB of combined GPU memory. Thinking Machines lists configurations including four Nvidia B300 GPUs or eight Nvidia H200 GPUs. A compressed NVFP4 version reduces the memory requirement to about 180 GB, but still targets professional hardware.

The smaller footprint may nevertheless matter to companies running their own AI infrastructure. It can reduce inference costs and make fine-tuning or private deployment more accessible than with the much larger Inkling model.

Thinking Machines says it improved the model through changes to training data and training methods. It also used an earlier Inkling-Small checkpoint for on-policy distillation, with Inkling serving as a teacher model, before continuing reinforcement learning focused on agentic coding.

The release adds to a growing market for open-weight models that organisations can adapt for internal tools, content workflows and specialised applications. Its practical value will depend less on aggregate benchmark scores than on performance in a company’s own tasks, available hardware and requirements for accuracy, privacy and governance.

Sources

• Introducing Inkling Small – Thinking Machines Lab Blog
• Thinking Machines debuts Inkling Small open source AI model nearing performance of predecessor at about 1/4 size – VentureBeat