Alibaba

A runtime for the execution layer under GPU kernel agents, from Wei Wang's group at HKUST (Tianyu Feng and Haoxuan Yu first authors) with seven co-aut…

Alibaba's Accio team argues that most steps in everyday agent work are state tracking, coordination, recovery, and follow-through rather than frontier…

Qwen's production speech-recognition system, documented in a 45-author technical report on 7 September 2026: a Mixture-of-Experts, LLM-based ASR model…

"Turning Agent Trajectories into Scalable Terminal Environments." The Qwen Team, with Tsinghua, observes that terminal-agent trajectories now exist at…

"Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents." Alibaba Research diagnoses two failur…

"Stop Training Agents on Contexts They Never Saw." Alibaba's Token Foundry (Pengjun Xie, Jieping Ye) observes that harnesses which compress context mi…

"Evaluating LLM Agents on Long-Horizon Autonomous Business Operations." A Qwen Team benchmark, built with Taobao and Tmall and HKUST, in which an agen…

The Qwen Team (with Huazhong University of Science and Technology) extends the Qwen line into autonomous driving: Qwen-Drive-1.0 keeps the natively mu…

Alibaba with Zhejiang University observes that production agent skills are directory bundles, not isolated prompts: the root loads at activation while…

Open-weight preview of the Qwen4 architecture, playing the role Qwen3-Next played for Qwen3.5: a multimodal (image/video-in, text-out) MoE with a 125B…

An open robotic value-foundation-model family from Alibaba DAMO Academy and Hupan Lab. It learns temporal distance to a language-specified goal from m…

Alibaba's first open-weight Qwen-Max-class release and its largest public model: a 2.4T-parameter MoE with 95B active per token. Its 92-layer hybrid s…

Tongyi Lab's 14B end-to-end character-animation DiT consumes a driving video directly, avoiding lossy pose or motion extractors. A dual branch, time-a…

A Qwen × HKU XLANG Lab collaboration (first author Dunjie Lu; Qwen authors include Shuai Bai): a native computer-use agent on a 397B-A17B Qwen M…

Names and measures Solution Hacking: frontier models reaching correct answers on science benchmarks through invalid shortcuts (numerical search, enume…

Trains VLMs to use visual tools as a compositional and adaptive capability rather than fixed-tool-space grounding: a hierarchical trajectory-synthesis…

QwenLM's answer to text-only agent harnesses: native multimodal plugins that "make any agent harness multimodal-native", wiring Qwen's vision/audio mo…

Multimodal embedder (2B/4B/9B, Qwen3.5 backbones) emitting dense + SPLADE-style sparse embeddings in one causal pass — text, image, and video, i…

DAMO Academy's medical MLLM (8B/32B, Apache 2.0) on Qwen3-VL-32B with a DINOv2 + CLIP-ConvNeXt cascade-fusion vision stack, native 2D and 3D NIfTI CT/…

DAMO Academy and Hupan Lab's embodied foundation-model family adds native 3D grounding, contact-point prediction, and an action model spanning Unitree…

Tongyi/Wan's ~25B single-stage diffusion long-form music generator: five-minute multilingual songs with dual-stem (vocals + BGM) output; claims 7.43%…

A 14B image-to-video character-animation model from Alibaba's Wan (Tongyi Wanxiang) video team, specialized for dance and full-body character motion &…

DAMO Academy's embodied world models, released as a pair (both HF-trending at launch). RynnWorld-4D predicts future RGB, depth, and optical flow from…

A universal audio-generation framework that unifies speech, sound, and music in a single model by extending autoregressive next-token prediction from…

Qwen's first "native language world model" for agents: given an agent action, it predicts the next environment state across seven agent domains (MCP t…

Qwen's entry into robotics: three simultaneous technical reports, with weights explicitly not released. RobotNav is a navigation foundation model with…

Pathology vision foundation model from DAMO Academy (Alibaba Group), CAS Institute of Automation, and Hupan Lab. ViT-L backbone (plus a ViT-B "DaX-Bas…

The cost tier of Alibaba's Qwen3.7 generation, sitting below the Qwen3.7-Max flagship and succeeding Qwen3.6-Plus. Proprietary and API-only (qwen3.7-p…

Alibaba's new flagship reasoning model, unveiled at the 2026 Alibaba Cloud Summit on May 20 after debuting anonymously on LM Arena around May 14. Prop…

Shows that full-attention LLMs are already intrinsically sparse and can be converted into highly sparse models with only a few hundred training steps…

"Think When Needed: Adaptive Reasoning-Driven Multimodal Embeddings with a Dual-LoRA Architecture." Proposes a dual-LoRA setup that attaches a reasoni…

Qwen Team with MBZUAI (and KAUST): a systematic study of structured pruning and knowledge distillation for MoE models at pretraining scale. Findings:…

Open-weight Qwen3.6 family released under Apache 2.0. Two flagships: Qwen3.6-27B (dense, 27B, 64 layers, hybrid gated DeltaNet + gated attention) achi…

Top proprietary tier of the Qwen3.6 generation and the first fully closed-weights Qwen Max flagship — released only as a Preview ("smarter, shar…

The Qwen3.5 generation of the Qwen-Omni line: an end-to-end any-to-text-and-speech model that, unlike its open-weight predecessors, ships API-only on…

Addresses when agentic multimodal models should use external tools versus relying on internal knowledge. Proposes HDPO (Hybrid Decoupled Policy Optimi…

Enterprise-focused frontier model with 1M token context (up to 65,536 output tokens). Always-on chain-of-thought reasoning and native function calling…

Introduces Think-Anywhere, enabling LLMs to invoke reasoning blocks at arbitrary token positions during code generation, rather than only before imple…

Addresses skill routing for LLM agents at scale (10K+ skills). Key finding: hiding the skill body causes a 31–44 percentage point drop in routin…

Fourth generation of the Qwen model family. Qwen3.5-397B-A17B native VLM built on Qwen3-Next architecture. 201 languages and dialects supported. 1M co…

Open web world model for training and evaluating browser agents: predicts next-page state given the current state and a candidate action. Three sizes…

Dataset for training reasoning reward models.

Qwen's open speech-recognition line (backfilled from January 2026; Transformers-native "-hf" ports shipped June 26): 0.6B and 1.7B models built on the…

Benchmark for evaluating complex agentic planning.

10-million-token benchmark for corpus-level analysis and reasoning — answers require global integration, comparison, and statistical aggregation acros…

Polymath learning asks how far reinforcement learning can go with a single, carefully engineered training problem. Starting from Qwen2.5-7B-Base, the…

A training-infrastructure system for multi-task agentic reinforcement learning on disaggregated hardware. Agentic RL workloads mix compute-bound prefi…

Safety-specialized guardrail model for real-time token stream moderation.

"Fin-PRM: A Domain-Specialized Process Reward Model for Financial Reasoning in Large Language Models." A process reward model purpose-built for financ…

Image generation model family. Qwen-Image (20B MMDiT) excels at text rendering in Chinese and English. Qwen-Image-2.0 (7B) unifies generation and edit…

CLI tool for terminal-based agentic coding and codebase analysis.

Third-generation coding-specialized model family including Qwen3-Coder-480B and Qwen3-Coder-Next.

Awarded **Best Paper at NeurIPS 2025**, this research introduces a foundational architectural modification for transformers: applying a head-specific…

Third generation of the Qwen model family. Introduced a massive 36-trillion token pretraining dataset. Includes dense, MoE, multimodal, speech, and re…

High-quality mathematical reasoning dataset.

End-to-end multimodal model processing text, images, audio, and video with real-time speech generation. Thinker-Talker architecture. Over 80k download…

Reasoning model following the chain-of-thought (CoT) paradigm.

Vision-language model with dynamic resolution processing and absolute time encoding.

Multimodal reasoning model following the chain-of-thought (CoT) paradigm.

Benchmark for evaluating the step-by-step logic of LLMs.

The "18-trillion token" series. Released sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B, plus a proprietary MoE model (Qwen2.5-Max).

Naive Dynamic Resolution for variable-resolution visual processing. Available in 2B, 8B, and 72B variants.

Coding-specialized model family. Code-specific models (0.5B-32B) pre-trained on 5.5T tokens.

Math-specialized model family. Includes the Qwen2.5-Math series (1.5B, 7B, 72B) which introduced Tool-Integrated Reasoning (TIR).

Significantly improved multilingual capabilities. Sizes: 0.5B, 1.5B, 7B, 72B, and the 57B-A14B MoE.

Major update (the "Qwen2 Beta") introducing 0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B, and 110B dense models, alongside a 14B MoE variant.

Audio model family. Evolution from Qwen-Audio (2023) to Qwen2-Audio-7B (Aug 2024).

Python framework for building agents with tool-calling, RAG, and code execution.

C++ implementation for efficient local inference of Qwen models.

First-generation vision-language model in the Qwen family.

First generation of the Qwen large language model series, spanning 1.8B to 72B parameters.

Contrastive vision-language pretraining specifically for Chinese. Predates the Qwen branding.

Unified sequence-to-sequence framework for multimodal tasks. Predates the Qwen branding.

Multi-modality-to-multi-modality mega-transformer, including the M6-10T (10 trillion parameters) and M6-T sparse models. Predates the Qwen branding.