發布日期2026/04
推論速度136 tokens/s
首字延遲時間0.99s
輸入價格$0.375/1M
輸出價格$2.25/1M
排名表現
評測分析
Qwen3.6 35B A3B is a solid choice for projects requiring multimodal processing and reasoning, but exhibits clear limitations in coding and natural conversation. It is best suited for AI and data teams seeking an economical model for multimodal workloads, rather than a coding assistant or conversational chatbot.
強項
- Robust multimodal performance ranking in the global top 10; handles images, text, and cross-modal interactions effectively
- Solid logical reasoning, suitable for complex analytical tasks and multi-step deduction
- Reliable AI agent performance, well-suited for low-to-medium complexity automation workflows
弱項
- Weak coding capabilities; unsuitable for complex code generation or debugging
- Subpar general conversational skills; responses can be disjointed or unnatural in standard dialogue
- Context window and throughput lag behind leading models, limiting its effectiveness for long-context or real-time tasks
適用情境
Multimodal document and image analysis (OCR, visual recognition, and data extraction from scans)Logical reasoning, complex deduction, and multi-step problem-solvingBasic workflow automation and AI agents for document routing and classification
指南與影片
Qwen3.6-35B-A3B is an MoE (Mixture-of-Experts) model from Alibaba featuring 35B total parameters with only 3B active per token. Access it directly via DeepInfra or Roboflow, or download it from Hugging Face for local deployment. It handles programming, code analysis, and multimodal processing (text, images, video). Tip: Use Q4_K_M quantization to run on a 24GB GPU, or deploy quickly via Ollama.