'Qwen3.8-27B,' a locally runnable version, has been released for free, and it surpasses the Claude Opus 4.6 Max in some benchmarks.



Alibaba's Qwen team has released ' Qwen3.8-27B, ' an AI model with 27 billion parameters. The model's weights are released under the Apache License 2.0, and commercial use is permitted under certain conditions. While it is an open-weight model that can be installed on any PC, some performance tests have reported it outperforms Anthropic's high-performance model, ' Claude Opus 4.6 Max .'




The Qwen3.8 series was announced for release in July 2026, with the 'Qwen3.8-Max-Preview' version, featuring 2.4 trillion parameters, initially becoming available. Alibaba then announced the official version, 'Qwen3.8-Max,' in August. Qwen3.8-Max is a model that emphasizes software development and long-term autonomous work, boasting capabilities such as working for over 125 hours to reproduce and improve research papers and continuing software development for 16 days without human intervention.

Alibaba announces the open model release of 'Qwen3.8,' a large-scale model with 2.4 trillion parameters, boasting 'performance second only to Claude Fable 5' - GIGAZINE



Furthermore, Alibaba has also released the model data for 'Qwen3.8-2.4T-A95B,' which served as the basis for Qwen3.8-Max, free of charge. Qwen3.8-2.4T-A95B boasts a massive 2.4 trillion parameters, but employs a 'Mixture-of-Experts (MoE)' method that only moves a portion of them depending on the processing task, meaning only 95 billion parameters are used during inference. Even so, a data center-level computing environment is required to run it.

Chinese AI 'Qwen3.8' with 2.4 trillion parameters released as an open model for free, boasting performance comparable to cutting-edge models from OpenAI and Anthropic - GIGAZINE



Even if a high-performance open model is released, it will be difficult for individuals or small development environments to run it if it requires enormous computing resources. That's why the Qwen team announced Qwen3.8-27B, which reduces the number of parameters to 27 billion, when they released Qwen3.8-Max.

Qwen3.8-27B is a Dense model that utilizes the entire model during inference, rather than a MoE model that only runs the necessary parts. The Qwen team describes it as a 'compact and easily deployable model.' The model data is released under the Apache License 2.0 and can be used with Hugging Face Transformers, as well as vLLM, SGLang, and Docker Model Runner, which are used to run AI models at high speed. An FP8 quantized version to reduce size and computational load is also officially available.

Qwen/Qwen3.8-27B · Hugging Face
https://huggingface.co/Qwen/Qwen3.8-27B



Qwen3.8-27B also offers improved performance compared to Qwen3.6-27B, which has the same 27 billion parameters. The following is an excerpt of the main benchmark results released by the Qwen team. The highest score among the three models listed is shown in bold. Qwen3.8-27B outperformed Claude Opus 4.6 Max in SWE-bench Pro, which measures software development capabilities, and CoWorkBench, which measures long-term work performance. On the other hand, Claude Opus 4.6 Max outperformed in Terminal Bench 2.1 and GPQA Diamond, which measures scientific inference capabilities.

Qwen3.8-27B Qwen3.6-27B Claude Opus 4.6 Max
Coding in the terminal
Terminal Bench 2.1		 73.0 63.4 78.2
Software Development
SWE-bench Pro		 61.7 53.5 53.4
Performing long hours of work
CoWorkBench		 70.7 61.0 68.2
Following instructions
IFBench		 79.5 69.1 62.5
Reasoning in the field of science
GPQA Diamond		 89.2 87.8 91.3



Qwen3.8-27B is a multimodal model that can handle not only text but also images and videos. Designed as a Vision-Language model that can handle images and videos as standard, it allows you to operate a PC or smartphone while viewing the screen, and to read documents and scientific diagrams.

It can understand long videos and can handle videos up to several hours long. The standard context length is 262,144 tokens, and with extension technology, it can support up to 1 million tokens.

The main multimodal-related benchmarks are as follows:

Qwen3.8-27B Qwen3.6-27B Claude Opus 4.6 Max
PC operation
OSWorld - Updated		 84.3 63.9 72.7
Smartphone operation
AndroidWorld		 81.9 70.3 62.0
Software development including images
SWE-MM		 38.6 25.7 27.1
Understanding diagrams and charts in the scientific field
CharXiv (RQ, Without CI)		 83.7 78.4 66.0
Document comprehension
OmniDocBench 1.5		 91.1 89.4 86.6



Support for local execution environments has been progressing since its release. LM Studio states that at least 17GB of RAM is required for the minimum quantization version, and Qwen3.8-27B can also be used with Ollama.




AMD has also reported that it will work with PCs equipped with Ryzen AI Max+ and Radeon AI PRO R9700 with 32GB of VRAM.




According to tests by SGLang, using a GeForce RTX 5090 and Qwen3.8-27B with optimizations such as NVFP4 applied, the decoding speed reached 206.1 tokens per second.




Qwen3.8-27B also supports Thinking Mode, which allows for inference before answering, and the amount of inference can be adjusted in three stages: 'xhigh,' 'medium,' and 'low.' The Qwen team plans to offer a hosted version of Qwen3.8-27B on Qwen Cloud in the future, which will include a context length of 1 million tokens and official built-in tools as standard.

Related Posts:

in AI, Posted by log1d_ts