Aug 18, 2026 11:48:00

Analysis shows that the locally running Chinese-made open model 'Qwen3.8 27B' has agent performance exceeding that of GPT-5.6 Terra, making it overwhelmingly more powerful than models of comparable size.

Chinese company Alibaba released the locally executable AI model '

Qwen3.8 27B ' on August 15, 2026 (Japan time). On August 17, 2026, the third-party organization Artificial Analysis published its performance analysis results for Qwen3.8 27B, revealing that despite being an open model that runs locally, Qwen3.8 27B possesses performance comparable to closed models such as GPT-5.6 Terra.

Qwen3.8 27B - Intelligence, Performance & Price Analysis
https://artificialanalysis.ai/models/qwen3-8-27b

The Qwen3.8 27B is an open-source model released free of charge, and can run on high-performance home PCs using AI execution applications such as Ollama and LM Studio. Benchmark results published by Alibaba revealed that it outperformed the Claude Opus 4.6 Max in several tests.

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

The following are the test results for Qwen3.8 27B using Artificial Analysis. The Artificial Analysis Intelligence Index score, which calculates intelligence performance by running multiple benchmark tests, was '52', matching that of GPT-5.6 Luna.

The Artificial Analysis Agentic Index score, which calculates agent performance, is '51,' surpassing GPT-5.6 Terra and Claude Opus 4.8.

The graph below shows the total number of model parameters on the horizontal axis and the Artificial Analysis Intelligence Index score on the vertical axis. It can be seen that Qwen3.8 27B is smaller in scale and runs locally compared to models with similar performance, yet it has high performance.

The graph below compares the number of output tokens per task. Qwen3.8 27B is characterized by its significantly higher output token count. For models used in APIs or subscription services, a large number of output tokens directly translates to higher costs, but with Qwen3.8 27B, it can be run locally, so there is no need to consider API charges, etc. On the other hand, problems such as 'output tokens accumulating and reaching the context window limit faster' and 'increased memory usage due to accumulated output tokens' occur.

Artificial Analysis has evaluated the performance of Qwen3.8 27B, stating that it is 'a top-class model in terms of intelligence performance' and that 'it output a total of 160 million tokens in the Artificial Analysis Intelligence Index test, which is extremely redundant compared to the median of 43 million tokens.'