NVIDIA releases NVFP4-quantized Qwen3.6-35B checkpoint on Hugging Face
4-bit NVFP4 build of Alibaba’s Qwen3.6-35B-A3B lands with Apache-2.0 license, long-context support, and vLLM instructions; NVIDIA notes it did not develop the base model.
By Ryan Merket · Published
Why it matters
Pre-quantized, permissively licensed checkpoints reduce the time and cost to get a capable long-context MoE model into production, especially for vLLM users on Hopper and Blackwell GPUs.
NVIDIA published a quantized version of Alibaba's Qwen3.6-35B-A3B model on Hugging Face on May 28, 2026, packaged as Qwen3.6-35B-A3B-NVFP4 under the Apache-2.0 license. NVIDIA stresses the underlying model is third-party and not developed by NVIDIA; this release applies NVIDIA's quantization for deployment.
The checkpoint targets production inference: a Mixture-of-Experts transformer with hybrid attention, listed at 35B parameters with 3B activated, and a context window up to 262K tokens. Inputs span text, image, and video, with text-only outputs. NVIDIA positions the build for AI agents, chatbots, RAG, and related applications.
Quantization was performed with NVIDIA Model Optimizer to the NVFP4 data type (version 1.0, modelopt v0.44.0). NVIDIA says it quantized weights and activations for linear operators in transformer MoE blocks, cutting precision from 16-bit to 4-bit and shrinking disk and GPU memory footprint by about 3.06x. Calibration used cnn_dailymail and NVIDIA's Nemotron-Post-Training-Dataset-v2. NVIDIA lists evaluation coverage across MMLU Pro, GPQA Diamond, tau^2-Bench Telecom, MMMU Pro, SciCode, AIME 2025, AA-LCR, and IFBench without reporting scores.
Deployment guidance centers on vLLM, with sample serve commands and DGX Spark environment variables provided on the model card. NVIDIA cites compatibility with Hopper and Blackwell GPUs and tested inference on an NVIDIA GB300. The release notes that additional use-case testing is required to meet safety and compliance needs.