Qwen3.8-27B - A new generation of large-scale AI model open-sourced by Alibaba's Qianwen team.

What is Qwen3.8-27B?

Qwen3.8-27B is a new generation of large-scale models open-sourced by Alibaba's Qianwen team. It employs a native multimodal dense architecture with 27 billion parameters, supports 262K native contexts, and can be extended to 1M tokens via YaRN. After quantization, it can be smoothly deployed on consumer-grade graphics cards. The model adds a new reasoning_effort function, which can dynamically adjust the depth of thought according to the task difficulty, demonstrating particularly outstanding performance in programming and agent capabilities: achieving a score of 61.7 on SWE-bench Pro and 84.3 on the OSWorld PC benchmark, both surpassing Qwen3.7-Plus and Claude Opus 4.6 Max.

Main functions of Qwen3.8-27B

• Native multimodal understandingIt supports end-to-end joint understanding of images, videos, and text, and can handle STEM diagrams, complex documents, and long video content.
• Long context processingIt natively supports 262K context windows and can extrapolate to 1M tokens through YaRN technology, easily handling long documents and codebase-level tasks.
• Intelligent thinking adjustment (reasoning_effort)The depth of thinking is dynamically controlled according to the difficulty of the task; simple tasks save resources, while complex tasks involve deep reasoning.
• Strong programming skillsIt supports agentic coding, terminal programming, and repository-level code generation. Its SWE-bench Pro score is 61.7, surpassing Qwen3.7-Plus.
• Multi-terminal agent operation (Agentic Use)It can control computers (OSWorld 84.3 points), browsers (WebArena 64.8 points), and mobile phones (AndroidWorld 81.9 points) to complete complex tasks end-to-end.
• Office and Long-Term Task ExecutionIt excels at long-term office tasks (CoWorkBench 70.7 points) and professional-level work tasks (JobBench 33.4 points), a significant improvement over the previous generation.
• Visual application developmentSupports application replication (RecreationBench 47.1 points) and visual web development (Vision2Web 62.9 points), and can generate interface code by viewing images.
• High-efficiency local deploymentIt features a dense architecture with 27 billion parameters, which, after quantization, can run smoothly on consumer-grade graphics cards and is compatible with mainstream inference frameworks such as Transformers, vLLM, and SGLang.
• Open commercial licensingIt is open source under the Apache 2.0 permissive license, supporting free download, deployment, and commercial use.

How to use Qwen3.8-27B

• Download model weightsDownload for free from Hugging Face or ModelScope on the Moda community. Moda is recommended for users in China as it offers faster download speeds.
• Select inference framework for deployment：
  • TransformersSuitable for rapid verification and development debugging.AutoModelForCausalLM / AutoProcessor Direct loading
  • vLLM / SGLangSuitable for high-concurrency inference in production environments, supporting tensor parallelism and continuous batch processing.
• Run locally with one clickQuantized versions (such as INT4) can run on a single consumer-grade graphics card (approximately 16–24GB of video memory), making them suitable for local deployment by individual developers.
• Enable thinking modeThe model defaults to thinking mode; this can be changed by setting... reasoning_effort Parameters (such as low / medium / high) control the depth of thinking; lower the parameter for simple questions and answers to save resources, and raise the parameter for complex reasoning to ensure quality.
• Extended long contextNative support for 262K; enable inference configuration when longer context is required. YaRN Parameters (such as vLLM) rope_scaling (Configuration) can be extrapolated to 1M Tokens.
• Invoking multimodal capabilities: Input images, videos, or documents as shown in the official examples. This can be used for tasks such as chart analysis, document understanding, and visual web page generation.
• Build an Agent applicationIt combines Claude Code-like programming frameworks, OSWorld-like computer control environments, or browser/mobile automation toolchains to leverage its agentic operational capabilities.
• Waiting for the official APIIf you do not want to deploy it yourself, Qianwen has announced that it will launch an API version, which will provide a context of 1 million tokens by default and can be directly called from the cloud.
• Adhere to open source licenseThe model is licensed under the Apache 2.0 license and can be used freely for commercial purposes, provided the copyright notice is retained.

Technical Principles of Qwen3.8-27B

• Dense architectureIt adopts a full activation design with 27 billion parameters, which is different from the MoE hybrid expert model. It activates all parameters in each inference, resulting in more stable behavior and simpler deployment.
• Hybrid attention mechanismIt adopts a hybrid structure that combines Gated DeltaNet (gated linear attention) and Gated Attention (gated full attention), with a total of 64 layers and 5120 hidden dimensions. Linear attention is responsible for efficiently processing long sequences, while full attention ensures accurate modeling of key information, balancing efficiency and performance.
• Native multimodal architectureThe vision and text capabilities are deeply integrated from the pre-training stage (rather than grafting a visual encoder later), enabling end-to-end image, video, and document understanding.
• Ultra-long context technologyNative support for 262K context, via YaRN(Position encoding extrapolation technique) can be scaled to 1M tokens without retraining, and the hybrid attention structure makes the computational cost of long sequence inference significantly lower than that of the pure full attention model.
• Adjustable thinking mechanism (reasoning_effort)The default setting enables the thinking mode, which allows the model to allocate computational load according to task difficulty through parameterized control of inference depth, thus achieving flexible scheduling of "test-time computation".
• Agentic skills trainingSpecialized post-training (including reinforcement learning) is performed on intelligent agent tasks such as programming, computer/browser/mobile phone operation to enable the model to have long-term planning, tool calling and multi-step execution capabilities.
• Consumer-grade deployment optimizationThe dense, small-size design, combined with quantization techniques (such as INT4/INT8), allows for smooth operation on a single consumer graphics card and is 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
• Magic Dash Community：https://www.modelscope.cn/collections/Qwen/Qwen38

Qwen3.8-27B's core advantages

• Small size, big capabilityWith a dense architecture of only 27 billion parameters, its overall performance surpasses that of the larger Qwen3.7-Plus from the previous generation, and even exceeds top closed-source models such as Claude Opus 4.6 Max in some benchmarks (such as SWE-bench Pro score of 61.7).
• Agent capabilities are among the world's leading.It has achieved top rankings in three major intelligent agent benchmarks: computer operation (OSWorld 84.3), browser operation (WebArena 64.8), and mobile phone operation (AndroidWorld 81.9), and can complete complex tasks end-to-end.
• Consumer-grade hardware can runAfter quantification, a single home graphics card can be deployed smoothly, allowing top-level agent capabilities to move from the cloud to individual developers and small and medium-sized enterprises.
• Extremely long contextWith a native 262K context, YaRN extrapolates to 1M tokens, easily handling long documents, large codebases, and hour-long videos.
• Native MultimodalIt performs comprehensively in multimodal tasks such as deep fusion training of images, videos, and text, chart analysis, document understanding, and visual web page development.
• Computing power is elastic and controllableThe reasoning_effort function dynamically adjusts the depth of thought based on the difficulty of the task, balancing performance and reasoning cost.
• Programming skills improved significantlyCompared to Qwen 3.6-27B, DeepSWE jumped from 13.3 points to 42.2 points, and QwenSWEBench jumped from 49.3 points to 79.0 points.
• Zero barriers to commercial useApache 2.0 Permissive License: Free to download, deploy, and use commercially without licensing restrictions.

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

Application scenarios of Qwen3.8-27B

• AI Programming AssistantIt serves as a local code generation, bug fixing, and repository-level refactoring assistant, scoring 61.7 on SWE-bench Pro, making it suitable for secure development of enterprise internal private code repositories.
• Computer-aided office automation (Computer Use)OSWorld scores 84.3, supporting real-world office scenarios with its automated desktop software operation, file organization, report filling, and cross-application data transfer capabilities.
• Browser-based Web AgentIt automatically completes browser tasks such as web page information retrieval, price comparison, form filling, and data collection.
• Mobile Use Automation: Controlling Android devices to complete tasks such as app testing, process automation, and accessibility assistance (AndroidWorld 81.9 points).
• Long document/long video analysisLeveraging the 1M context, it can digest entire financial reports, legal files, collections of papers, or hourly conference recordings in one go and output summaries and insights.
• Enterprise Knowledge Base Question and Answer (RAG)Local deployment ensures data remains within the internal network, and combined with long context, it directly "feeds" enterprise documents, making it suitable for highly compliant industries such as finance, healthcare, and law firms.