Thinking Machines releases its first open-weights model — Inkling has 975B parameters, 41B active
TL;DR
Thinking Machines released Inkling, a 975B open-weights model with 41B parameters active per token.
Thinking Machines Lab has released the first open-weights AI model it trained from scratch. Inkling, released July 15, is a 975B-parameter multimodal Mixture-of-Experts transformer with 41B parameters active per token and a context window of up to 1M tokens. The company also previewed Inkling-Small, which has 276B total parameters and 12B active.
Inkling uses a 66-layer decoder, with each token selecting six of 256 routed experts. Thinking Machines says pretraining used 45 trillion text, image, audio and video tokens, followed by more than 30 million reinforcement-learning rollouts on NVIDIA GB300 NVL72 systems.
The model processes text, images and audio with adjustable thinking effort. Company-reported scores are 77.6% on SWE-bench Verified, 97.1% on AIME 2026 and 73.5% on MMMU Pro; Thinking Machines says Inkling is not the strongest model overall.
The full weights are on Hugging Face under Apache 2.0. The BF16 checkpoint requires at least 2 TB of VRAM; NVFP4 lowers that to 600 GB and can run on four NVIDIA B300 GPUs. Tinker offers 64K and 256K context options.
via MarkTechPost / Thinking Machines Lab / Inkling Model Card / Hugging Face
Inkling uses a 66-layer decoder, with each token selecting six of 256 routed experts. Thinking Machines says pretraining used 45 trillion text, image, audio and video tokens, followed by more than 30 million reinforcement-learning rollouts on NVIDIA GB300 NVL72 systems.
The model processes text, images and audio with adjustable thinking effort. Company-reported scores are 77.6% on SWE-bench Verified, 97.1% on AIME 2026 and 73.5% on MMMU Pro; Thinking Machines says Inkling is not the strongest model overall.
The full weights are on Hugging Face under Apache 2.0. The BF16 checkpoint requires at least 2 TB of VRAM; NVFP4 lowers that to 600 GB and can run on four NVIDIA B300 GPUs. Tinker offers 64K and 256K context options.
via MarkTechPost / Thinking Machines Lab / Inkling Model Card / Hugging Face
