Thinking Machines releases Inkling, a 1T-parameter multimodal LLM

Thinking Machines has released Inkling, a large open-source multimodal language model available on Hugging Face that natively processes text, image, and audio inputs. The model features approximately 1 trillion total parameters with 41 billion active parameters, trained on 45 trillion tokens to support reasoning across modalities.

• Architecture: A decoder-only Mixture-of-Experts (MoE) model with 256 experts, utilizing relative attention, hybrid global/sliding window attention, and short 1D convolutions.
• Context and Inference: Supports a 1 million token context window and includes speculative MTP layers for faster inference; available in BF16 and NVFP4 variants.
• Deployment: Offers day-0 support for Hugging Face Transformers, SGLang, vLLM, and llama.cpp, with VRAM requirements of 2 TB for BF16 and 600 GB for NVFP4.

The model is intended for domain adaptation via fine-tuning to build multimodal reasoning applications.