Inkling-Small NVFP4 lands on Hugging Face for multimodal tuning

An open-weight checkpoint for multimodal work

One important signal in multimodal AI is not only a lab demo, but whether developers can download a checkpoint and tune it. On July 30, 2026, NVIDIA AI highlighted Thinking Machines’ Inkling-Small NVFP4 release. The tweet said: Another open-weight release from @thinkymachines. Inkling-Small is here. With native reasoning over audio and images and variable thinking effort, it is a great choice for fine-tuning, with NVIDIA NeMo on NVIDIA DGX Station. The source post is available on X.

The linked Hugging Face repository adds useful detail. thinkingmachines/Inkling-Small-NVFP4 is listed as a quantized version of thinkingmachines/Inkling-Small under the Apache-2.0 license. Its tags include transformers, safetensors, image-text-to-text, audio-text-to-text, conversational, and MoE. Hugging Face API metadata shows the NVFP4 checkpoint was created on July 27, modified on July 30, and had passed 25,000 downloads at the time of review.

NVIDIA AI’s account often points to model releases that fit its GPU, NeMo, DGX, inference, and training stack. That makes the tweet more practical than promotional: NVIDIA is positioning Inkling-Small as a candidate for fine-tuning on NeMo and DGX Station. The NVFP4 checkpoint also fits a broader trend in which lower-precision variants reduce deployment and experimentation cost while preserving enough capability for task-specific adaptation.

The next thing to watch is independent evaluation. Open weights and an Apache-2.0 license make the model easier to test, but they do not prove quality, safety, or reliability. Developers should compare the base Inkling-Small model with the NVFP4 checkpoint, measure degradation after quantization, and run their own checks on long conversations, image-and-audio prompts, tool use, and domain-specific fine-tuning before using it in production services.