Qwen3.5 122B A10B

Qwen3.5-122B A10B is a Mixture-of-Experts model designed to balance massive capability with deployment practicality. It carries 122 billion total parameters but activates only 10 billion per forward pass, using a hybrid Gated DeltaNet and MoE architecture where linear attention layers and full attention are interleaved in a 3-to-1 ratio for efficient long-context processing. Early fusion multimodal training gives it cross-generational parity with its text-only siblings across reasoning, coding, agents, and visual understanding benchmarks, and its vision performance even surpasses the larger Qwen3-VL-235B. As a mid-tier option in the Qwen3.5 family, it sits between the flagship 397B model and smaller variants, offering a practical sweet spot for teams that need frontier-level performance without flagship-level resource demands.

The model benefits from reinforcement learning scaled across million-agent environments with progressively complex task distributions, which the developers describe as supporting robust real-world adaptability. Alongside broad framework compatibility—Hugging Face Transformers, vLLM, SGLang, and KTransformers—it supports 201 languages and dialects for global deployment. Hardware benchmarks show meaningful throughput across different configurations, reaching over 2,000 tokens per second on 4× B200 GPUs with vLLM, and competitive performance on H100 and A100 setups as well. Available as open weights on Hugging Face, it serves developers and enterprises seeking a high-capability vision-language model that remains efficient enough for production use cases.