Thinking Machines Lab has released Inkling-Small, an open-weights Mixture-of-Experts model with 276 billion total parameters and 12 billion active parameters. The model is trained to reason natively over text, images, and audio, featuring a 1M token context window and adjustable thinking effort.

  • Weights are available under Apache 2.0 on Hugging Face, with BF16 requiring 600 GB VRAM and NVFP4 quantization dropping the requirement to 180 GB.
  • The architecture uses 42 decoder layers where each token routes to 6 of 256 experts plus 2 shared experts, processing images via patches and audio as dMel spectrograms.
  • Inkling-Small surpasses its larger teacher model Inkling on reasoning benchmarks like Humanity's Last Exam (31.6% vs 29.7%) and SWE-bench Verified (80.2% vs 77.6%).
  • Factual recall regressed significantly, with SimpleQA Verified dropping to 20.6% from Inkling's 43.9%, while multimodal scores remained close to the larger model.

The release enables startups and enterprises to self-host a frontier-capable model on single or dual GPU setups, supporting workloads like coding agents, terminal automation, and call-center analytics.