General Q&A · Coding
Qwen3.6-35B-NVFP4
A 35B multimodal mixture-of-experts model deployable on one MegaCube, using NVIDIA NVFP4 quantization for general Q&A, document understanding, coding, and agent tool calling.
01 / Model overview
Model architecture & capabilities
Qwen3.6-35B-A3B is Qwen's multimodal mixture-of-experts model, activating approx. 3B parameters at each step. It is trained for Q&A, writing, document understanding, coding, and agent tasks, including multi-turn tool use that retains reasoning context. NVIDIA's NVFP4 version uses low-precision quantization to reduce weight storage and compute requirements.
- Publisher / quantizer
- Alibaba Qwen / NVIDIA
- Full model name
- nvidia/Qwen3.6-35B-A3B-NVFP4
- Parameter count
- 35B total parameters; approx. 3B active per token
- Architecture
- 40-layer hybrid-attention MoE; 256 routed experts, 8 selected per step, plus 1 shared expert
- Model context
- Native 262,144 tokens, expandable to 1,010,000 tokens; actual service length depends on runtime settings
- Input / output
- Text, image, and video input; text and tool-call output
- Weight precision
- NVIDIA ModelOpt NVFP4
- License
- Apache-2.0
02 / Use cases
Use cases
Everyday Q&A and information synthesis
Ask questions about work materials, extract key points, and write summaries to turn long documents into clear conclusions.
Code generation, debugging, and editing
Write code from requirements, explain existing logic, analyze errors, and help make code changes.
Image understanding and tool use
Understand images and charts, analyze them together with text, and complete multi-step tasks through tool calling.
Plan model deployment and application integration
Evaluate device configurations and inference services for your model size, workloads, and concurrency needs.