Alibaba releases Qwen 3.8-Max, a 2.4 trillion parameter model (95B active) with improvements across coding, work, research, and long-horizon tasks. Open weights will be released next week.
Alibaba launches Qwen-Image-3.0, a third-generation foundational image generation model that supports up to 4.5k token input, precise rendering of text as small as 10px, native rendering of 12…
Alibaba releases Qwen-AgentWorld, a native language world model for simulating agent environments across seven domains, along with the AgentWorldBench benchmark.
Alibaba announces the Qwen-Robot Suite, a set of three foundation models (Qwen-RobotNav, Qwen-RobotManip, Qwen-RobotWorld) that bridge vision-language understanding with physical action in…
Alibaba announces Qwen-RobotManip, a generalizable Vision-Language-Action (VLA) foundation model built upon Qwen-VL, which introduces a unified alignment framework across representation, motion, and…
Alibaba introduces Qwen-RobotWorld, a dual-stream diffusion world model that uses natural language as a universal action interface to unify 20+ robot embodiments and 500+ action categories, enabling…
Qwen3.7-Plus is a multimodal agent model that unifies vision and language, capable of operating as a multimodal interactive hybrid agent, coding agent, productivity assistant, and visual agent, with…
Alibaba announces Qwen-VLA, a general-purpose Vision-Language-Action model built on the Qwen multimodal backbone, designed to extend visual perception, language understanding, and spatial reasoning…
Alibaba introduces Qwen3.7-Max, a proprietary model designed for the agent era, with strong capabilities in coding, general-purpose agents, reasoning, and multilingual tasks, as demonstrated by…
Alibaba announces Qwen3.5-LiveTranslate-Flash, a real-time multimodal translation model with expanded language coverage, ultra-low latency, voice cloning, and hotword enhancement, built on the…
Alibaba introduces Qwen-Scope, an interpretability toolkit that uses Sparse Autoencoders (SAEs) trained on Qwen3 and Qwen3.5 series models to decompose hidden representations into interpretable…
Alibaba's Qwen team open-sourced FlashQLA, a high-performance linear attention kernel library for GDN (Gated Delta Network) that achieves 2-3x forward and 2x backward speedup over the FLA Triton…
Alibaba open-sources Qwen3.6-27B, a dense 27-billion-parameter multimodal model that supports both thinking and non-thinking modes and delivers flagship-level agentic coding performance, surpassing…
Alibaba's Qwen team releases Qwen3.6-Max-Preview, an early preview of their next proprietary model with improved agentic coding, world knowledge, and instruction following over Qwen3.6-Plus.
Alibaba open-sourced the Qwen3.6-35B-A3B model, a sparse mixture-of-experts model with 35 billion total parameters and 3 billion active parameters, demonstrating strong agentic coding performance and…
Alibaba announces Qwen3.6-Plus, a new model with enhanced agentic coding capabilities, multimodal reasoning, and a 1M context window, available immediately via API.
Alibaba releases Qwen3.5, an open-weight native vision-language model with 397B total parameters (17B activated), featuring hybrid linear attention and sparse MoE architecture, a 1M token context…
Alibaba launched Qwen-Image-2.0, a next-generation image generation model that supports 1k-token instructions for professional infographics, 2K resolution for realistic scenes, unified generation and…
Qwen3-Coder-Next is an open-weight language model designed for coding agents and local development, built on Qwen3-Next-80B-A3B-Base with hybrid attention and MoE.
Alibaba's Qwen team open-sourced the Qwen3-ASR family: two ASR models (1.7B and 0.6B parameters) and a forced alignment model (0.6B), all under Apache 2.0 license.
Alibaba's Qwen team announces Qwen3-Max-Thinking, a new flagship reasoning model with adaptive tool-use capabilities and a test-time scaling strategy, now available in Qwen Chat and via API.
Alibaba's Qwen team has open-sourced the Qwen3-TTS family of speech generation models, including two sizes (1.7B and 0.6B parameters), supporting voice clone, voice design, and natural language-based…
Alibaba releases Qwen3-VL-Embedding and Qwen3-VL-Reranker, new multimodal retrieval models built on Qwen3-VL, supporting text, images, screenshots, and video.
Alibaba introduces Qwen-Image-2512, the December update of its Qwen-Image text-to-image foundational model, with key improvements in human realism, finer natural detail, and improved text rendering.
Alibaba released Qwen-Image-Edit-2511, an updated image editing model with improved character consistency, multi-person consistency, built-in LoRA integration, and enhanced geometric reasoning.
Alibaba's Qwen3-TTS family launches two new models: a voice design model (Qwen3-TTS-VD-Flash) and a voice cloning model (Qwen3-TTS-VC-Flash), both accessible via the Qwen API.
Alibaba introduces Qwen-Image-Layered, a model that decomposes images into multiple RGBA layers for independent editing, supporting variable-layer and recursive decomposition.
Qwen3-Omni-Flash-2025-12-01 is an upgraded version of Qwen3-Omni, a native multimodal model that processes text, images, audio, and video, and outputs text and speech simultaneously.
A new reinforcement learning method called SAPO (Soft Adaptive Policy Optimization) is introduced for training large language models, replacing hard clipping with a smooth, temperature-controlled…
Alibaba updated the Qwen3-TTS model family, including Qwen3-TTS-Flash, with richer timbres (over 49), support for 10 languages and 9 dialects, and improved prosody and speech rate naturalness.
Alibaba's Qwen announces an updated version of its deep research tool, Qwen DeepResearch 2511, featuring a multi-agent collaborative mechanism, improved anti-hallucination, report quality, and new…
Alibaba announces Qwen3Guard, the first safety guardrail model in the Qwen family, built on Qwen3 foundation models and fine-tuned for safety classification.
Alibaba introduces Qwen-Image-Edit, an image editing model built upon the 20B Qwen-Image model, extending text rendering capabilities to image editing for precise text editing, and using Qwen2.5-VL…
Alibaba proposes the Group Sequence Policy Optimization (GSPO) algorithm to address instability and model collapse issues in existing reinforcement learning algorithms for language models, aiming to…
Alibaba introduces an update to Qwen-MT (qwen-mt-turbo) via Qwen API, built on Qwen3 with trillions of multilingual and translation tokens, and reinforcement learning to improve translation accuracy…
Alibaba's Qwen team announces Qwen3-Coder, a new agentic code model, with its most powerful variant Qwen3-Coder-480B-A35B-Instruct (480B parameters, 35B active, MoE) supporting 256K native context…
Qwen-TTS now supports generating speech in three Chinese dialects: Pekingese, Shanghainese, and Sichuanese, and offers seven Chinese-English bilingual voices.
Alibaba introduces Qwen VLo, a unified multimodal understanding and generation model that can both understand images and generate high-quality recreations.
Alibaba announces the release of the Qwen3 family of large language models, including Qwen3-235B-A22B, Qwen3-30B-A3B, and Qwen3-4B, with competitive benchmark results.
Alibaba's Qwen team officially released the first version of QVQ-Max, a visual reasoning model that can understand and reason with content from images and videos.
Alibaba releases Qwen2.5-Omni, a new end-to-end multimodal model that processes text, images, audio, and video and generates real-time streaming responses via text and speech.
Alibaba's Qwen team announced the open-source release of Qwen2.5-VL-32B-Instruct, a 32B parameter vision-language model under Apache 2.0, refined from the Qwen2.5-VL series using reinforcement…
The post discusses the benefits of scaling data and model size, notes limited experience in scaling large models, references DeepSeek V3, and states that Alibaba is developing Qwen2.
Alibaba's Qwen team released two open-source models, Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, which support a context length of up to 1 million tokens.