Qwen: Qwen3.5-122B-A10B
Model Type
Proprietary Model
API access only
Recommended Use Cases
Try Qwen3.5-122B-A10B
Qwen3.5-122B-A10B is a high-capacity MoE model with 122B total parameters and 10B active, designed for long-horizon agentic tasks requiring sustained logical consistency.
Overview
Released February 24, 2026, Qwen3.5-122B-A10B sits at the high end of the open-weight Qwen3.5 lineup. With 10 billion active parameters, it maintains logical consistency across extended reasoning chains and massive context windows, narrowing the gap between open-weight models and proprietary frontier systems for complex agentic workloads.
Key Features
- 122B total / 10B active parameters (MoE)
- 256K native context (extendable to 1M+ with YaRN)
- Long-horizon optimization: Tuned for multi-step agentic tasks
- High needle-in-haystack accuracy across full context window
- Native multimodal: Text, image, and video understanding
- 201 languages supported
- Apache 2.0 license
Benchmark Highlights
- AIME 2026: 85% success rate (top tier for reasoning models)
- Strong performance on long-document analysis and codebase understanding
- Approaches proprietary frontier model capability
When to Use Qwen3.5-122B-A10B
Choose Qwen3.5-122B-A10B when you need:
- Complex long-horizon agentic workflows
- Analysis of large documents (500+ page contracts, massive codebases)
- Maximum open-weight capability
- Sustained reasoning over extended context
- Tasks requiring frontier-level intelligence
Choose Qwen3.5-35B-A3B when you need:
- Consumer hardware deployment
- Faster inference with lower resource requirements
- Cost-effective local deployment
Choose Qwen3.5-27B when you need:
- Dense model stability
- Better quantization tolerance
- Simpler deployment without MoE complexity
Choose Qwen3.5-Plus (API) when you need:
- Maximum capability without self-hosting
- Adaptive tool use
Hardware Requirements
| Quantization | VRAM Required |
|---|---|
| 4-bit (MXFP4) | ~80GB |
| 8-bit | ~140GB |
| FP16 | ~250GB |
Designed for server-grade GPUs (80GB+ VRAM). Full weight matrix must remain in memory even with sparse activation.
Role in Series
Qwen3.5 medium models (Feb 24, 2026):
- Qwen3.5-122B-A10B: Maximum capability, server deployment (this model)
- Qwen3.5-35B-A3B: Best efficiency, consumer hardware
- Qwen3.5-27B: Dense stability, easier quantization