Thinking Machines launches Inkling-Small, matching its flagship at a quarter the size

Thinking Machines Lab on Thursday released Inkling-Small, a multimodal open-weight model that matches or exceeds its predecessor Inkling while using roughly one-quarter of the parameters — a move that dramatically lowers the compute requirements for running a frontier-class AI system.

A Smaller Model That Punches Up

Inkling-Small uses a Mixture-of-Experts architecture with 276 billion total parameters and 12 billion activated per inference, compared to Inkling's 975 billion total and 41 billion active parameters. The model accepts text, image, and audio inputs and supports a context window of up to one million tokens.siliconrepublic+1

According to Thinking Machines, the model matches or exceeds Inkling on reasoning and agentic coding tasks. On Humanity's Last Exam, Inkling-Small scored above 31 percent compared to Inkling's 29.7 percent, and on SWE-Bench Verified it crossed 80 percent. Artificial Analysis scored it 40 on its Intelligence Index, just one point below Inkling's 41.artificialanalysis+1

The company achieved these results through a two-step post-training process: on-policy distillation using Inkling as a teacher model, followed by two weeks of agent-focused reinforcement learning.kucoin

Lowering the Barrier to Entry

Where the original Inkling required a minimum of two terabytes of aggregated VRAM to run BF16 checkpoints, Inkling-Small brings those requirements down sharply. The model is available in both BF16 and NVFP4 formats on Hugging Face and through Thinking Machines' Tinker service.gigazine+1

"Tinker customers have seen first-hand that the right fine-tuned model can outperform closed models on a variety of tasks, and do so faster and cheaper," the company said. The reduced scale makes LoRA and full-parameter fine-tuning feasible for mid-sized teams — a capability previously limited to large corporations.kucoin+1

Turbulence Behind the Scenes

The release comes amid ongoing leadership upheaval at the Nvidia ↘ 1.47%-backed startup. Lilian Weng, one of six original co-founders, announced her departure on July 26 citing health reasons, with her last day on July 29. She has since rejoined OpenAI to lead work on recursive self-improvement. Only Mira Murati and John Schulman now remain from the founding team.kucoin+3

Thinking Machines launched its first model, Inkling, on July 15. Less than two weeks later, it shipped a successor that outperforms it — a pace that suggests, as the company framed it, a "production line" for generating models rather than a one-off effort.kucoin+1