Atomic-Germ/Qwen3.6-35B-A3B-NPU2

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Atomic-Germ/Qwen3.6-35B-A3B-NPU2 is a 35.1 billion parameter Mixture-of-Experts (MoE) model, based on the Qwen3.6 architecture, specifically converted to the Q4NX packed-quantization format for FastFlowLM. This variant is optimized for AMD Ryzen AI XDNA2 (NPU2) processors and includes both text and vision Q4NX weights for multimodal input. It requires approximately 51 GB of unified system memory and is designed for high-performance inference on compatible NPU hardware.

Loading preview...

Atomic-Germ/Qwen3.6-35B-A3B-NPU2 Overview

This model is a specialized conversion of the Qwen3.6-35B-A3B Mixture-of-Experts (MoE) architecture, developed by Atomic-Germ. It is specifically designed for deployment on AMD Ryzen AI XDNA2 (NPU2) processors, such as those found in Strix Point / Ryzen AI 300 series.

Key Features and Optimizations

  • Q4NX Quantization: Utilizes FastFlowLM's native Q4NX packed-quantization format, a rearranged Q4_1 layout tuned for NPU matrix engine's tile sizes and memory access patterns. This format is exclusive to FastFlowLM and does not run on llama.cpp or Ollama.
  • Multimodal Capabilities: Ships with both text and vision Q4NX weights, enabling multimodal input processing.
  • Hardware Specific: Requires an AMD Ryzen AI processor with XDNA2 (NPU2) and the XRT NPU stack installed on Linux.
  • Memory Footprint: Demands approximately 51 GB of unified system memory for weights, activations, and KV cache.
  • Context Length: Supports a substantial context length of 262,144 tokens.

Usage and Installation

Installation is managed via the flm-add installer, which registers the model with FastFlowLM without modifying the system-wide installation. Users need FastFlowLM version 0.9.45 or later. The model leverages NPU kernels from the official Qwen3.6-35B-A3B-NPU2 model due to shared engine family and architecture.

When to Use This Model

This model is ideal for developers and researchers who:

  • Possess AMD Ryzen AI XDNA2 (NPU2) hardware.
  • Require high-performance, on-device inference for large language models.
  • Are working with multimodal applications that benefit from combined text and vision processing.
  • Are already integrated into the FastFlowLM ecosystem and seek optimized Qwen3.6 performance on their specific hardware.

Finetunes
1 models