OpenAI spin-offs at Thinking Machines cut Inkling’s compute bill by three-quarters with their new Inkling-Small model. It’s a Mixture-of-Experts transformer with 276 billion total parameters but only 12 billion active at inference. Trained on NVIDIA GB300 NVL72 clusters, it handles text, images and audio natively, stretches to a 1 million-token context window and matches full-size Inkling on benchmarks like agentic tool use (Terminal-Bench 2.1), text-only reasoning (HLE) and instruction following (IFBench). Its “thinking effort” knob lets you dial performance up or down to hit a cost-performance sweet spot. They tweaked the data mix and training recipe based on lessons from Inkling. After a preview checkpoint, they ran two weeks of on-policy distillation with Inkling as teacher and more agentic coding RL. The result: Inkling-Small pulls ahead on reasoning and coding benchmarks—31.6 percent on Humanity’s Last Exam versus Inkling’s 29.7 percent—and holds that edge across all compute budgets. In multimodal tests, it nearly equals Inkling on visual reasoning, chart and diagram understanding, math Q&A with images, speech tasks and long-form audio reasoning, all at lower cost. It processes images in 40×40-pixel patches via a four-layer hMLP and audio as dMel spectrograms, then blends these with text tokens. Safety and forecasting stayed top of mind. Inkling-Small reuses Inkling’s safety recipe and red-teaming results. On the FORTRESS adversarial test it refuses harmful queries 71.6 percent of the time (Inkling hits 78.0 percent), and it accepts benign ones 96.9 percent. On StrongREJECT it scores 98.4 percent refusal on unambiguously harmful prompts. In forecasting (ForecastBench, Prophet Arena) it posts Brier indices around 61.3 for no-search, ticking above Inkling’s 60.1. The full weights are open-sourced, fine-tuning is available on Tinker, and you can try text, image and audio chat in the Tinker Playground.