Digital Event Horizon
Think you know what's possible with language models? Welcome Inkling by Thinking Machines is about to change everything with its revolutionary multimodal capabilities.
The Welcome Inkling model is a revolutionary large language model developed by Thinking Machines that can accept image and audio inputs in addition to text. The model uses a decoder-only architecture with a mixture-of-experts structure for efficient processing of large amounts of data. The model is trained on an enormous dataset of 45 trillion tokens, images, audio, and video. The model's capabilities are enhanced by its agentic nature, allowing it to reason across modalities. The model comes in two variants with different VRAM requirements and supports remote inference and deployment. The model has been demonstrated to excel in various use cases such as agentic coding, multimodal vision, and multimodal audio.
Welcome Inkling is a revolutionary large language model developed by Thinking Machines, which has been recently released on Hugging Face. This groundbreaking model is designed to natively accept image and audio inputs in addition to text, making it a game-changer in the field of multimodal natural language processing.
The Welcome Inkling model is built around a decoder-only architecture with a mixture-of-experts (MoE) structure that utilizes relative attention and hybrid attention schemes to efficiently process large amounts of data. The model's transformer-based encoder consists of 256 experts, each of which can be activated based on the router scores. This allows for faster inference and more efficient processing of complex queries.
Inkling is trained on an enormous dataset of 45 trillion tokens, images, audio, and video, making it one of the largest open models available today. The model's capabilities are further enhanced by its agentic nature, which enables it to reason across modalities such as audio, images, and text. This makes Welcome Inkling an ideal solution for building new wave multimodal reasoning applications.
The model comes in two variants: full BF16 and a well-calibrated NVFP4 variant. The BF16 variant requires 2 TB of VRAM, while the NVFP4 variant requires only 600 GB of VRAM. Additionally, the model supports remote inference with Hugging Face Inference Providers, local inference using llama.cpp and Unsloth, and can be deployed on a cluster using SLURM scripts.
Welcome Inkling has been demonstrated to excel in various use cases such as agentic coding with Pi, multi-token prediction drafters, multimodal vision, and multimodal audio. The model's performance is further evaluated in benchmarking results against other large language models. The dataset mentioned includes ArtificialAnalysis/big_bench_audio, iNLP-Lab/GlobeAudio, and others.
In conclusion, Welcome Inkling by Thinking Machines is a groundbreaking multimodal large language model that has the potential to revolutionize natural language processing applications. Its unique architecture, extensive training data, and capabilities make it an ideal solution for building new wave multimodal reasoning applications.
Related Information:
https://www.digitaleventhorizon.com/articles/Welcome-Inkling-by-Thinking-Machines-A-Revolutionary-Multimodal-Large-Language-Model-deh.shtml
https://huggingface.co/blog/thinkingmachines-inkling
Published: Wed Jul 15 13:52:47 2026 by llama3.2 3B Q4_K_M
