Qwen3.8-27B - A New Generation AI Large Model Open-Sourced by the Alibaba Qwen Team

What is Qwen3.8-27B

Qwen3.8-27B is a new generation large model open-sourced by the Alibaba Qwen team. It adopts a native multimodal dense architecture with 27 billion parameters, supports a native 262K context and can be extended to 1M Tokens via YaRN, and after quantization can be smoothly deployed on consumer-grade GPUs. The model adds a reasoning_effort feature that can dynamically adjust the depth of thinking according to task difficulty. It performs especially prominently in coding and Agent capabilities: SWE-bench Pro reaches 61.7 points, and the OSWorld computer operation benchmark reaches 84.3 points, both surpassing Qwen3.7-Plus and Claude Opus 4.6 Max.

Main Features of Qwen3.8-27B

• Native Multimodal Understanding: End-to-end support for joint understanding of images, videos, and text, and can handle STEM charts, complex documents, and long video content.
• Ultra-Long Context Processing: Natively supports a 262K context window and can be extrapolated to 1M Tokens via YaRN technology, easily handling long documents and codebase-level tasks.
• Intelligent Thinking Adjustment (reasoning_effort): Dynamically controls the depth of thinking according to task difficulty, saving resources on simple tasks and performing deep reasoning on complex tasks.
• Powerful Coding Capabilities: Supports Agentic Coding, terminal coding, and repository-level code generation, with a SWE-bench Pro score of 61.7, surpassing Qwen3.7-Plus.
• Multi-Platform Agentic Use: Can operate computers (OSWorld 84.3 points), browsers (WebArena 64.8 points), and phones (AndroidWorld 81.9 points), completing complex tasks end to end.
• Office and Long-Horizon Task Execution: Excels at long-cycle office tasks (CoWorkBench 70.7 points) and professional-level work tasks (JobBench 33.4 points), a substantial improvement over the previous generation.
• Visual Application Development: Supports application recreation (RecreationBench 47.1 points) and visual web development (Vision2Web 62.9 points); interface code can be generated just by looking at an image.
• Efficient Local Deployment: With a 27-billion-parameter dense architecture, it can run smoothly on consumer-grade GPUs after quantization and is compatible with mainstream inference frameworks such as Transformers, vLLM, and SGLang.
• Open Commercial Licensing: Open-sourced under the permissive Apache 2.0 license, supporting free download, deployment, and commercial use.

How to Use Qwen3.8-27B

• Download Model Weights: Download for free from Hugging Face or the ModelScope community; domestic users are recommended to use ModelScope for faster speeds.
• Choose an Inference Framework for Deployment：
  • Transformers: Suitable for quick validation and development debugging,AutoModelForCausalLM / AutoProcessor Directly load
  • vLLM / SGLang: Suitable for high-concurrency inference in production environments, supporting tensor parallelism and continuous batching
• One-Click Local Run: Quantized versions (such as INT4) can run on a single consumer-grade GPU (about 16–24GB VRAM), suitable for local deployment by individual developers.
• Enable Thinking Mode: The model enables thinking mode by default; by setting the reasoning_effort parameter (such as low / medium / high), you can control the depth of thinking. Lower it for simple Q&A to save resources, and raise it for complex reasoning to ensure quality.
• Extend Ultra-Long Context: Natively supports 262K; when a longer context is needed, enable the YaRN parameter in the inference configuration (such as vLLM's rope_scaling configuration) to extrapolate to 1M Tokens.
• Call Multimodal Capabilities: Feed in images, videos, or document inputs as in the official examples; it can be used for tasks such as chart analysis, document understanding, and visual web page generation.
• Build Agent Applications: Combine it with Claude Code-style programming frameworks, OSWorld-style computer control environments, or browser/mobile automation toolchains to leverage its Agentic operation capabilities.
• Wait for the Official API: If you do not want to deploy it yourself, Qwen's official team has already announced that an API version will be released, offering a 1 million token context by default and allowing direct cloud calls.
• Comply with the Open Source License: The model uses the Apache 2.0 license, allowing free commercial use as long as the copyright notice is retained.

Technical Principles of Qwen3.8-27B

• Dense Architecture: It uses a fully activated design with 27 billion parameters. Unlike MoE mixture-of-experts models, it activates all parameters in every inference, resulting in more stable behavior and simpler deployment.
• Hybrid Attention Mechanism: It uses a hybrid structure combining Gated DeltaNet (gated linear attention) and Gated Attention (gated full attention), with 64 layers in total and a hidden dimension of 5120. Linear attention is responsible for efficiently handling long sequences, while full attention ensures precise modeling of key information, balancing efficiency and performance.
• Native Multimodal Architecture: Vision and text capabilities are deeply integrated from the pretraining stage (rather than attaching a vision encoder later), enabling end-to-end understanding of images, videos, and documents.
• Ultra-Long Context Technology: It natively supports a 262K context, and through YaRN(a positional encoding extrapolation technique), it can be extended to 1M tokens without retraining. The hybrid attention structure makes the computational overhead of long-sequence inference significantly lower than that of pure full-attention models.
• Adjustable Reasoning Mechanism (reasoning_effort): Thinking mode is enabled by default and controlled through reasoning depth parameterization, allowing the model to allocate computation according to task difficulty and enabling elastic scheduling of "Test-time Compute".
• Agentic Capability Training: Specialized post-training (including reinforcement learning) is conducted for agent tasks such as programming and computer/browser/mobile operation, equipping the model with long-horizon planning, tool calling, and multi-step execution capabilities.
• Consumer-Grade Deployment Optimization: The dense small-size design, combined with quantization techniques (such as INT4/INT8), allows it to run smoothly on a single home GPU and remain compatible with mainstream inference frameworks such as Transformers, vLLM, and SGLang.

Project Address of Qwen3.8-27B

• Hugging Face：https://huggingface.co/collections/Qwen/qwen38
• ModelScope Community：https://www.modelscope.cn/collections/Qwen/Qwen38

Core Advantages of Qwen3.8-27B

• Small Size, Large Capability: With only a 27-billion-parameter dense architecture, its overall performance surpasses the larger previous-generation Qwen3.7-Plus, and on some benchmarks (such as SWE-bench Pro 61.7) it even exceeds top closed-source models such as Claude Opus 4.6 Max.
• World-Leading Agent Capabilities: It tops all three major agent benchmarks: computer operation (OSWorld 84.3), browser operation (WebArena 64.8), and mobile operation (AndroidWorld 81.9), and can complete complex tasks end to end.
• Runs on Consumer-Grade Hardware: After quantization, it can be deployed smoothly on a single home GPU, bringing top-tier Agent capabilities from the cloud to individual developers and small and medium-sized enterprises.
• Ultra-Long Context: Native 262K context, extrapolated with YaRN to 1M tokens, easily handling long documents, large codebases, and hour-level long videos.
• Native Multimodal: Deep fusion training of images, video, and text, with comprehensive performance in multimodal tasks such as chart analysis, document understanding, and visual web development.
• Elastic and controllable compute: The reasoning_effort feature dynamically adjusts thinking depth based on task difficulty, balancing performance and reasoning cost.
• Major leap in coding capability: Compared with Qwen3.6-27B, DeepSWE jumped from 13.3 points to 42.2 points, and QwenSWEBench jumped from 49.3 points to 79.0 points
• Zero barrier to commercial use: Apache 2.0 permissive license, free to download, deploy, and use commercially, with no licensing restrictions.

Comparison of Qwen3.8-27B with similar competing products

Application scenarios of Qwen3.8-27B

• AI coding assistant: Locally build an assistant for code generation, bug fixing, and repository-level refactoring, with SWE-bench Pro score of 61.7, suitable for secure development of enterprise internal private code repositories.
• Computer automation office work (Computer Use): Automatically operate desktop software, organize files, fill out reports, and transfer data across applications, with OSWorld score of 84.3 supporting implementation in real office scenarios.
• Browser agent (Web Agent): Automatically complete browser tasks such as web information retrieval, price comparison, form filling, and data collection.
• Mobile automation (Mobile Use): Control Android devices to complete tasks such as app testing, process automation, and accessibility assistance (AndroidWorld score of 81.9).
• Long document/long video analysis: Leveraging a 1M context, it can digest an entire financial report, legal case file, paper collection, or hour-long meeting recording at once and output summaries and insights.
• Enterprise knowledge base question answering (RAG): Local deployment ensures data does not leave the intranet, and combined with long context it can directly "feed" enterprise documents, suitable for highly compliant industries such as finance, healthcare, and law firms.