Thinking Machines Lab Introduces Inkling-Small: 276B-Parameter MoE Model Now Runnable on Single GPU

Thinking Machines Lab has released Inkling-Small, an open-weights Mixture-of-Experts model with 276 billion total parameters and 12 billion active, roughly one-quarter the size of Inkling. Trained on NVIDIA GB300 NVL72 systems, the model natively reasons over text, images, and audio with a 1-million-token context window, distributed under Apache 2.0. The critical advantage is computational accessibility. The BF16 checkpoint requires at minimum 600GB of aggregated VRAM, achievable with 4 B300 or 8 H200 GPUs. Quantized versions substantially lower this ceiling: the NVFP4 checkpoint runs on 180GB, enabling single-GPU B300 inference in W4A4. This opens a 276-billion-parameter model to startups renting cloud instances and mid-size enterprises with existing H200 capacity. On reasoning and coding benchmarks, Inkling-Small exceeds its teacher model: 31.6% on Humanity's Last Exam versus Inkling's 29.7%, and 80.2% on SWE-bench Verified versus 77.6%. The model is encoder-free and natively multimodal, processing images as 40×40-pixel patches and audio as dMel spectrograms.