update
May 1, 2026
By Teun
NVIDIA posts NVFP4-quantized Gemma 4 26B model on Hugging Face
NVIDIA has released Gemma 4 26B IT NVFP4, a quantized version of Google DeepMind’s open multimodal model, on Hugging Face. The model supports text and image input, long-context work up to 256K tokens, and is designed for reasoning, coding, chat, and multimodal tasks on NVIDIA GPU systems.
NVIDIA has published Gemma 4 26B IT NVFP4 on Hugging Face, a quantized version of Google DeepMind’s Gemma 4 26B IT model that is aimed at running on NVIDIA GPU-accelerated systems. According to the model card, the release is ready for commercial and non-commercial use under Apache License 2.0 and is available globally.
The model is positioned as an open multimodal system for text and image tasks, with video support described as processing video as sequences of frames. NVIDIA says it is designed for chat, summarization, image data extraction, reasoning, coding, function calling, research, and education.
Gemma 4 26B IT NVFP4 is the NVIDIA Model Optimizer version of the model, with weights and activations quantized to the NVFP4 data type using the company’s nvfp4_experts_only recipe. NVIDIA says the goal is to make the checkpoint ready for inference with vLLM, the open-source inference server commonly used to serve large language models.
The model card says the release uses a hybrid attention design that mixes local sliding-window attention with full global attention. It also uses unified Keys and Values in global layers and Proportional RoPE, or p-RoPE, to support long-context performance.
The model has a 256K-token context window, which means it can hold and process a very large amount of text in a single request. NVIDIA says it supports over 140 languages, variable image aspect ratios and resolutions, and configurable visual token budgets of 70, 140, 280, 560, and 1120.
The architecture is listed as a 26B A4B MoE model with 25.2 billion total parameters and 3.8 billion active parameters. The model card says it has 30 layers, 8 active experts out of 128 total experts, plus one shared expert, and a 1,024-token sliding window.
For deployment, NVIDIA says the checkpoint is intended for Linux and NVIDIA Blackwell hardware, with vLLM as the supported runtime engine. The sample command in the model card includes tool-call and reasoning parsers for Gemma 4, plus auto tool choice and trust-remote-code flags.
NVIDIA also published benchmark results comparing the NVFP4 version against the full-precision baseline. On GPQA Diamond, the model scored 79.90% versus 80.30% for baseline. On AIME 2025 it scored 90.00% versus 88.95%, while on MMLU Pro it posted 84.80% versus 85.00%.
Other reported results were 79.80% on LiveCodeBench pass@1 versus 80.50% for baseline, 78.1% on IFBench versus 77.77%, and 96.40% on IFEval versus 96.60%. NVIDIA says the model was evaluated on text-based reasoning and coding benchmarks, including MMLU Pro, LiveCodeBench V6, SciCode, IFEval, GPQA Diamond, AIME 2025, and IFBench.
The company says the model was calibrated using cnn_dailymail and Nemotron-Post-Training-Dataset-v2, and that the underlying Gemma 4 26B IT model was trained on large-scale multimodal data spanning web documents, code, images, and audio with a cutoff date of January 2025. NVIDIA adds that the base model may reflect toxic language and societal biases from internet data, and that outputs may still be inaccurate, redundant, or undesirable.
At release time, NVIDIA noted an implementation limit: vLLM works with tensor parallelism set to 1 only, though expert parallelism is supported. The company said there is an open issue in vLLM and FlashInfer, and that the current MoE backend is VLLM_CUTLASS or Marlin, with a FlashInfer-TRTLLM path still under an open pull request.
