Thinking Machines releases Inkling as a huge open multimodal model

Thinking Machines Lab has released Inkling on Hugging Face, positioning it as a large open multimodal model for teams that want to adapt frontier-scale systems outside a closed API.

Inkling is described as a sparse mixture-of-experts model with roughly 975B total parameters and 41B active parameters. It can take text, image and audio inputs, supports a 1M-token context window, and is aimed at reasoning-heavy multimodal applications and domain fine-tuning. Hugging Face says the model is available in BF16 and NVFP4 variants, with support across Transformers, SGLang, vLLM and llama.cpp.

The launch also includes Inkling-Small, a 276B-parameter version with 12B active parameters, intended to make the architecture more practical to serve. Together AI is offering day-zero hosted access, which matters because the full model still needs hundreds of gigabytes of GPU memory even in lower-precision form.

The practical takeaway: another serious lab is pushing open-weight multimodal models toward long-context agents, voice/image workflows and custom enterprise fine-tuning, not just text chat.