⚡ BREAKING
Thinking Machines Releases Open Multimodal Model Inkling with 1 Trillion Parameters
Thinking Machines has released Inkling, an open multimodal LLM with ~1T total parameters and 41B active parameters, supporting text, image, and audio inputs with a 1M context window. The model is available on Hugging Face in BF16 and NVFP4 variants, with day-0 support in transformers, SGLang, vLLM, and llama.cpp.
Thinking Machines has open-sourced Inkling, a large multimodal mixture-of-experts model with approximately 1 trillion total parameters and 41 billion active parameters. It is the first large open model to natively accept text, image, and audio inputs with a 1M context window. The architecture includes relative attention instead of RoPE, hybrid attention with a 5:1 sliding-to-global ratio, short 1D convolution layers, and Mixture-of-Experts with shared experts. For vision, it uses a hierarchical MLP patchifier; for audio, it employs a discretized mel spectrogram. Inkling is available in full BF16 (requiring 2 TB VRAM) and a quantized NVFP4 variant (requiring 600 GB VRAM). It supports inference via transformers pipeline, AutoModelForMultimodalLM, SGLang, vLLM, and llama.cpp, and can be accessed through Hugging Face Inference Providers.
- Сокращения
- BF16 = Brain Floating Point 16-bit — 16-разрядный формат с плавающей запятой
- NVFP4 = NVIDIA Floating Point 4-bit — 4-разрядный формат с плавающей запятой NVIDIA
- MTP = Multi-Token Prediction — многотокеновое предсказание
- MoE = Mixture of Experts — смесь экспертов
- MLP = Multilayer Perceptron — многослойный перцептрон
- VRAM = Video Random Access Memory — видеопамять
- GPU = Graphics Processing Unit — графический процессор
Source: Hugging Face blog —
original