Agents A1 NVFP4 This repository contains an NVIDIA ModelOpt NVFP4 quantization of InternScience/Agents A1 , a 35B Qwen3.5 MoE agentic model. Credits and Attribution This NVFP4 checkpoint is derived from InternScience/Agents A1 . Base model: InternScience, for the Agents A1 model, training recipe, technical report, and original BF16 Hugging Face release. Quantization tooling: NVIDIA, for NVIDIA TensorRT Model Optimizer / NVIDIA ModelOpt , used to produce the NVFP4 ModelOpt checkpoint. Model architecture and runtime ecosystem: Hugging Face Transformers, Safetensors, Accelerate, and the Hugging Face Hub. Calibration data: CNN/DailyMail via Hugging Face Datasets, used for text path post training calibration. Inference ecosystem: vLLM/SGLang compatibility is inherited from the Qwen3.5 MoE / ModelOpt NVFP4 ecosystem, subject to runtime support and validation. Quantization Summary Field Value Base model InternScience/Agents A1 Quantization tool NVIDIA ModelOpt 0.44.0 Quantization format NVFP4 / ModelOpt FP4 ModelOpt config mtq.NVFP4 MLP ONLY CFG Calibration data abisee/cnn dailymail , text only calibration Calibration sequence length 1024 Architecture Qwen3 5MoeForConditionalGeneration Li…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy