Qwen3.5 122B A10B (NovitaAI)

Alibaba Cloud released Qwen3.5-122B-A10B on February 23, 2026 as a mid-tier multimodal foundation model under Apache 2.0 licensing, featuring 122B total parameters with 10B activated through a 256-expert MoE architecture. The model delivers strong results on MMLU-Pro (86.7%), GPQA Diamond (85.5%), SWE-bench Verified (72.4%), and Terminal-Bench 2.0 (41.6%) per aggregator data. Technical specifications include 262K native context extensible to 1M, grouped-query attention with 32 heads, 2 KV heads at head dimension 256, RoPE positional embedding, and SwiGLU FFN. Self-hosting requires substantial VRAM, such as three RTX 6000 Blackwell GPUs (288GB total) or one MI325X (256GB). Aggregator-listed API pricing stands at $0.40 input and $3.20 output per million tokens, reflecting serving-provider rather than creator pricing.

Qwen3.5-122B-A10B is a native vision-language model built on a hybrid architecture combining linear attention with a sparse mixture-of-experts design for higher inference efficiency. Released February 25, 2026, it features 122B parameters, a 262K context window, and multimodal input support including video, images, and text. It ranks second among Qwen3.5 variants in overall performance per aggregated benchmarks. The model shows an 88% reliability rate across eight benchmarks and excels in hallucination resistance (98.0%, 63rd percentile) and instruction following (75.0%, 80th percentile), with its text capabilities surpassing Qwen3-235B-2507 and visual capabilities exceeding Qwen3-VL-235B. However, it scores 0.0% in general knowledge, coding, reasoning, ethics, and mathematics tests, indicating clear improvement areas. Aggregator-reported provider pricing includes $0.40 input and $3.20 output per million tokens on Novita.