Thinking Machines Releases Inkling-Small with 276B Parameters, Performance Nears Inkling

Thinking Machines released Inkling-Small on July 31, 2026, a Mixture-of-Experts model with a total of 276 billion parameters and 12 billion activated per token. Its performance is comparable to the larger Inkling, but with significant optimizations in size and speed. Full weights are available on Hugging Face and Tinker Playground, supporting multimodal fine-tuning for text, image, and audio.

Fact Reconstruction

According to the company's official blog, Inkling-Small adopts a Mixture-of-Experts architecture with 276 billion total parameters and only 12 billion activated per token. It achieves performance close to Inkling on benchmarks such as Terminal-Bench 2.1, HLE text reasoning, and IFBench instruction following, through a variable thinking effort mechanism, while significantly reducing computational demands. Inkling itself has 975 billion total parameters and 41 billion activated, released on July 15, 2026.

Mechanism Breakdown

Inkling-Small's operation relies on the MoE design, which only activates a subset of expert networks to process input. This allows the 276-billion-parameter model to consume computing power equivalent to only 12 billion parameters during inference. The company trained the model on NVIDIA GB300 NVL72 systems and provides both the original and NVFP4 quantized versions for deployment on Blackwell hardware. The variable thinking effort feature allows users to balance output quality and cost across a range from minimal to xhigh, directly corresponding to output TFLOPs consumption.

Industry Impact

For developers, the full open weights mean the model can be directly downloaded and fine-tuned on the Tinker platform for domain-specific tasks without relying on API calls. For enterprise users, the design of 276 billion total parameters with only 12 billion activated lowers deployment barriers, especially suitable for scenarios requiring local inference. Regarding the competitive landscape, Thinking Machines' move provides an open-source option from the U.S. alongside Chinese open-source models like Qwen, emphasizing that organizations can shape the model themselves rather than relying on closed systems.

The official blog notes that Inkling-Small is more efficient than Inkling on benchmarks for agentic tool use, reasoning, and instruction following, and is competitive with open-source models of similar scale.

Comparison and Precedents

Compared to Inkling, Inkling-Small reduces total parameters from 975 billion to 276 billion, and activated parameters from 41 billion to 12 billion, while maintaining native multimodal support and a maximum context of 1 million tokens. Both models display a performance-compute curve on the same benchmarks, with Inkling-Small's curve shifted left, indicating lower cost at equivalent performance. Bridgewater Associates previously used similar open models to fine-tune financial tasks, reducing costs to one-fourteenth of those of closed models, a case that validates the practical value of open weights.

Strategic Assessment

Based on available facts, more organizations are likely to fine-tune Inkling-Small on the Tinker platform for specific workflows, to verify its actual cost-benefit under low and high effort settings.

When selecting a model, if tasks are latency-sensitive and require multimodal input, Inkling-Small's 12 billion activated parameters offer clear efficiency advantages. Enterprises with proprietary data can use the open weights to perform local fine-tuning, avoiding repeated payments to closed models.