Hugging Face valuation and Nvidia acquisition talks
Hugging Face maintains a $4.5 billion valuation from its 2023 funding round. On September 3, 2026, CEO Clement Delangue approached Jensen Huang to purchase the company for $12.93 billion. The company previously rejected a $500 million Nvidia investment that offered a $7 billion valuation to preserve its independence. Hugging Face reported $150 million in revenue in August 2026. The company employs 769 people as of May 2026 and hosts more than 1 million models. This platform is the infrastructure layer for the open-source AI ecosystem. It is the main distribution network for researchers and enterprises. Hugging Face hosts models from Meta’s Llama family and Mistral, providing a central library for open-source AI.
SmolVLM efficiency and robotic control
The SmolVLM family prioritizes memory-efficient multimodal inference through aggressive token compression and deployment-aware memory usage. The SmolVLM-256M variant requires less than one gigabyte of GPU memory, yet it outperforms the Idefics 80B system in AI2D science diagram benchmarks despite being significantly smaller in its total and comprehensive parameter count. The 500M variant achieves 90% of the performance seen in the 2.2B parameter model on key benchmarks. SmolVLA uses a Flow Matching action head to output robot joint commands from image observations. This SmolVLA model is capable of latency under 100ms while requiring roughly two gigabytes of VRAM. The 2.2B model version requires 5.5 gigabytes of VRAM at FP16 precision. For high-resolution images, SmolVLM uses an image-splitting strategy. SmolVLM2-2.2B is vulnerable in adverse weather conditions. When testing on the SVT cropped benchmark, the 2.2B model accuracy dropped from 86.86% to 28.59% under heavy rain and fog.
SmolVLM2 also provides Video2LoRA for video understanding. This method is a way to treat video understanding as parametric internalization instead of repeated video-in-context processing. It uses a frozen SmolVLM2 encoder and a trainable Perceiver hypernetwork to read intermediate hidden states. This approach reduces the answer-time visual-token load by up to 1,500x. The SmolVLM2-2.2B model uses Continuous Temporal Decoding for temporal grounding. This paradigm reaches a Charades-STA mIoU of 46.6.
MobileVLM V2 uses 1.21.2 million high-quality image-text pairs from ShareGPT4V to align vision-language features. It incorporates academic tasks like ScienceQA, TextVQA, SBU, and IConQA to increase instruction-following capacity. The training uses data from Docmatix to improve the model’s ability to handle document scans and detailed captions. The model uses a lightweight projector to reduce the number of visual tokens. This reduces the visual token load by up to 1,500x.
| Model | Parameters | VRAM (FP16) | Primary Use Case |
|---|---|---|---|
| SmolVLM 256M | 256M | ~1.2 GB | Real-time edge |
| SmolVLM 500M | 500M | ~1.5 GB | Lightweight OCR |
| SmolVLM 2.2B | 2.2B | ~5.5 GB | Document QA |
| SmolVLA | ~450M | ~2 GB | Robotic control |
Apple FastVLM and mobile deployment
Apple released FastVLM and MobileCLIP on Hugging Face to enable real-time VLM applications on mobile hardware. FastVLM uses the FastViT-HD hybrid encoder to generate fewer but higher-quality tokens. This design makes the model up to 85x faster and 3.4x smaller than previous architectures. FastVLM avoids the latency bottlenecks that higher image resolutions typically cause in traditional vision models. MobileCLIP-S2 is 2.3x faster and more accurate than ViT-B/16 baselines. These models use the MLX framework for machine learning on Apple Silicon. You should verify the repository name before downloading to avoid using older versions. The FastVLM demo is functional in real-time using transformers.js and WebGPU in a browser.
FastVLM is a hybrid architecture that includes a convolutional stem, three convolutional stages, and two transformer blocks. This structure is efficient during high-resolution processing. The model also uses a simple Multi-Layer Perceptron module to project visual tokens to the embedding space of the LLM. FastVLM is 5.2x faster than SmolVLM with a 0.5B LLM. It is also 21x faster than Cambrian-1 with a 7B LLM. Can these compact models eventually match the reasoning depth of much larger proprietary systems?
