source: hugging face blog: welcome inkling by thinking machines
level: technical
thinking machines released inkling, a decoder-only multimodal mixture-of-experts model with 975 billion total parameters and 41 billion active parameters. it accepts text, audio, and image inputs natively, using a simple hierarchical mlp patchifier for vision and a discretized mel spectrogram for audio. the architecture includes relative attention instead of rope, hybrid global and sliding window attention, and short 1d convolutions over hidden states. the mixture-of-experts layer routes tokens to 6 experts plus 2 shared experts always active.
inkling supports a 1 million token context window and was trained on 45 trillion tokens of text, images, audio, and video. it includes multi-token prediction drafters for speculative decoding, speeding up inference without changing outputs. the model is available in full bf16 requiring 2 tb of vram and a quantized nvfp4 version needing 600 gb. day-0 support is provided in transformers, sglang, vllm, and llama.cpp, with unsloth offering 1-bit quantized versions that reduce vram use by 95%.
the model is designed for reasoning across modalities and domain adaptation through fine-tuning. it can be used for agentic coding tasks with tools like pi, and supports serverless inference via hugging face inference providers. example code shows how to use the any-to-any pipeline or auto classes for text, image, and audio inputs. the release includes a drafter model for faster generation, and the community can run quantized versions locally on limited hardware.
why it matters: this model makes large-scale multimodal reasoning accessible for fine-tuning and building applications that combine text, images, and audio.
source: hugging face blog: welcome inkling by thinking machines