Alibaba's research lab behind the Qwen model family. Develops open-source LLMs, VLMs, and MoE architectures spanning 0.5B to 235B+ parameters, with dense and sparse (A3B) variants. Also produces Tongyi DeepResearch, a 30B-A3B MoE agentic LLM trained via GRPO-based RL on synthetic agentic trajectories for long-horizon web research tasks.
Researchers from Alibaba Group show that hybridizing attention at the head level — guided by mechanistic interpretability — unlocks 512K context with only 15B training tokens
Z-Image-Engineer V6 is a fine-tuned text encoder that rewrites simple prompts into cinematic descriptions and slots directly into Z-Image-Turbo workflows
Alibaba's Tongyi Lab ships a transparency-focused overhaul of its agent framework with observable events, retry logic, granular permissions, and pluggable workspaces.
Alibaba's Qwen3.7-Plus unifies screen perception, GUI control, and code generation into a single agent loop at one-sixth the cost of its text-only sibling
Alibaba's Qwen team unveils a single 5B model that controls 11 robot platforms across manipulation, navigation, and trajectory prediction with one set of weights.
Alibaba's Fun-Realtime-AudioChat and Fun-ASR claim three #1 spots on Artificial Analysis's speech leaderboard, beating GPT-Realtime and Gemini on reasoning and conversational dynamics.