Thinking Machines, the AI lab founded by Mira Murati, has released Inkling-Small — an open-weights model designed to deliver Inkling-level performance at a quarter of the size.
Inkling-Small packs 276 billion total parameters with 12 billion active per token, compared with 975 billion (41 billion active) for the flagship Inkling released in mid-July. The new model matches or beats its larger sibling on reasoning and agentic benchmarks: 31.6% vs 29.7% on Humanity's Last Exam, and above 80% on SWE-bench Verified. The company attributes the gains to an improved data mix, on-policy distillation from Inkling, and two extra weeks of reinforcement learning focused on agentic coding.
Like Inkling, Inkling-Small is natively multimodal — text, images and audio — with a 1M-token context and adjustable reasoning effort. The full weights are on Hugging Face, and the model is available for fine-tuning on the lab's Tinker platform.
The release lands amid a wave of efficient open-weight models — DeepSeek's V4-Flash-0731 update arrived the next day with a similar story — underscoring how fast the price-performance frontier is moving in the open-source AI market.
