The latest AI news and updates of Qwen — Alibaba's family of open-source and commercial models spanning text, vision, and code. Covering Qwen's latest launches, analysis, and product updates from the past 90 days.
QwenSep 3Qwen launched E-Commerce Bench, a benchmark evaluating AI agents on 365-day autonomous business operations using real market data. Agents manage inventory, pricing, and supplier negotiations across 6,886 products. The evaluation scores performance across seven axes, revealing that most models struggle to improve strategies over the year, with no single model dominating across all dimensions.
QwenSep 2Alibaba released Qwen3.8-Max-0902, an upgraded model featuring 2.4T parameters and a 1M-token context window. Post-trained on coding and collaborative work, the model debuted at #1 on the Code Arena: WebDev leaderboard with a score of 1,691. It is available via API on QwenCloud, priced at $2 per million input tokens and $6 per million output tokens.
QwenAug 27Alibaba’s experimental Qwen3.8-Flash-Next model now features day-0 support for NVIDIA NeMo AutoModel and NeMo RL, enabling domain-specific fine-tuning. The integration includes recipes for running the model with SGLang, vLLM, and LightSeek TokenSpeed. This release provides an early preview of the upcoming Qwen4 architecture, allowing immediate testing and optimization on NVIDIA hardware.
QwenAug 26Qwen released the Qwen3.8-Flash API on QwenCloud, priced at $0.16 per 1M input tokens and $0.47 per 1M output tokens. This multimodal mixture-of-experts model features a 125B parameter backbone with 6B active parameters and a 1M-token context window. It supports function calling, structured outputs, and context caching for high-concurrency agentic workflows.
QwenAug 26Alibaba released open weights for Qwen3.8-Flash-Next, a multimodal MoE model previewing the Qwen4 architecture. It features a 125B parameter backbone with 6B active parameters, utilizing GDN and QSA hybrid attention for efficiency. The model supports 262K native context, extensible to 1M tokens, and will arrive on QwenCloud at $0.16 per 1M input tokens and $0.47 per 1M output tokens.
QwenAug 22Alibaba Qwen’s Qwen3.8-27B is now the #1 open-weight model on Harvey’s Legal Agent benchmark, achieving 11.3 points. It ties with Fable 5 and outperforms Kimi K3, Qwen 3.8 Max, and DeepSeek V4. This 27-billion-parameter model delivers professional-grade agentic performance while remaining small enough to run on local hardware.
QwenAug 22Qwen amplified a community comparison showing its 27B open-weight model outperforming Google’s Gemini 3.7 Flash on a one-shot water-surface simulation. The model, running locally on an RTX 3090 at Q5 quantization, produced visually consistent results from a single prompt. Qwen now invites users to run the model on their own laptops.
QwenAug 15Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter model with 95 billion active parameters and a 1-million-token context window. The model is now available for dedicated inference through Day 0 launch partners Together AI, Fireworks AI, and Nebius Token Factory. It supports complex autonomous coding, research, and long-horizon agentic tasks.
QwenAug 10Alibaba released Qwen-MM-Plugins, an Apache-2.0 repository of skills and MCP servers that turn existing agent harnesses into multimodal-native systems. The plugins enable agents to read images, videos, and documents, edit video, and perform 3D or CAD modeling. A guided installer supports integration with harnesses like Claude Code and Codex, allowing agents to handle complex visual and spatial tasks.
QwenAug 6Alibaba's Qwen3.8-Max model now ranks fifth on the Artificial Analysis Intelligence Index and first on the Agentic Index. These independent evaluations measure the 2.4-trillion-parameter model's performance across nine diverse benchmarks, including coding and long-horizon tasks. The rankings confirm the model's top-tier agentic capabilities among 594 evaluated systems.
QwenAug 5Alibaba launched Qwen-Image-3.0-Pro on Qwen Cloud, offering high-fidelity image generation with support for 4.5k-token prompts and text rendering as small as 10px. The model supports 12 languages and 100+ art styles, with pricing starting at $0.04 per image for 1K resolution. It currently ranks second among mainstream models on the Text-to-Image Arena leaderboard.
QwenAug 3Alibaba reached the second position on the Text Arena leaderboard. The proprietary qwen3.8-max model achieved a preliminary score of 1496±10, placing the company behind only the top-ranked lab. This ranking reflects the model's performance across math, coding, and creative writing tasks.
QwenJul 31Alibaba released Qwen-Audio-3.0-ASR-Flash, an automatic speech recognition model featuring improved context consistency, custom hotword support, and structured transcript generation. Internal tests show the model achieves 95.36% recall on medical terminology and 93.24% on industrial terms. The model is available in streaming, file-based, and offline modes through Alibaba Cloud’s Model Studio.
QwenJul 22Qwen launched Qwen-Image-3.0, a foundational image generation model focused on realistic rendering. It supports 4.5k-token prompts for complex layouts like newspapers and nested UIs, renders text as small as 10px, and provides native support for 12 languages. The model also integrates live web retrieval to generate images based on current world knowledge and specific artistic styles.
QwenJul 19Alibaba announces Qwen3.8, a 2.4-trillion-parameter model slated for an upcoming open-weight release. The Qwen3.8-Max-Preview version is available immediately through Alibaba’s Token Plan, Qoder, and QoderWork platforms. The company characterizes the model as a frontier-level system, ranking it second only to Fable 5 in capability.
QwenJun 24Qwen released Qwen-AgentWorld, a native language world model simulating seven agent environments including Terminal, Web, and Android. The model uses a three-stage training pipeline to predict environment states, enabling controllable simulation for agent reinforcement learning. Qwen also released AgentWorldBench, where the 397B model outperforms GPT-5.4 and Claude Opus 4.8. The 35B-A3B version is now open-source.
QwenJun 16Alibaba released the Qwen-Robot Suite, a three-model stack for embodied intelligence. Qwen-RobotNav unifies five navigation tasks, Qwen-RobotManip aligns cross-embodiment manipulation using 38,100 hours of open-source data, and Qwen-RobotWorld predicts physically grounded futures across 20+ embodiments. These models function as composable physical-world tools for agentic systems, with a browser-based Chat2Robot demo available for testing manipulation capabilities.
Frequently asked questions