Qwen3.5 122B A10B

• 🧠 Mixture-of-Experts: 122B total parameters, ~10B active per token
• 🆕 Unified vision-language model handling text, image, and video
• 🔧 Hybrid Gated DeltaNet plus MoE architecture
• 📏 Natively supports long context up to 256K tokens
• 🎯 Reasoning, coding, agents, and visual understanding in one model
• 🔒 Runs in a Trusted Execution Environment with hardware attestation
• 🌐 Apache-2.0 licensed and multilingual
• 🏢 Built by Alibaba's Qwen team, served here privately by Venice

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen3.5 122B A10B is the mid-tier member of Alibaba's Qwen3.5 family, sitting between the flagship Qwen 3.5 397B and the lighter Qwen 3.5 35B A3B. It is a multimodal Mixture-of-Experts model with 122 billion total parameters but only about 10 billion activated per token, giving large-model capability at a much smaller inference cost. This Venice deployment additionally runs the model inside a Trusted Execution Environment, exposing hardware attestation so the enclave's identity and configuration can be independently verified.

Architecturally, the model uses a hybrid Gated DeltaNet plus MoE design that mixes linear-attention and full-attention layers, an approach Qwen positions for efficient long-context processing up to roughly 256K tokens. It natively accepts text, image, and video inputs and targets reasoning, coding, tool use, agentic workflows, and visual understanding.

Compared with its same-family predecessors, Qwen3.5 folds the previously separate vision line, such as Qwen3 VL 235B, into a single early-fusion multimodal foundation rather than keeping language and vision models distinct.

Within Venice's catalog it is succeeded by newer entries like Qwen 3.6 35B A3B FP8. Note that the underlying weights advertise a 256K context, while this listing is configured for 128K.