In early August, China’s Alibaba launched its updated flagship artificial intelligence model Qwen3.8-Max, which has 2.4 trillion parameters, 95 billion of which are still active. The model was designed to be used on cloud infrastructure and was initially only available for testing in QwenCloud, but the company later released its weights in the public domain for everyone. after she she released Qwen3.8-27B 27 billion parameter size, suitable for running on powerful PCs.
Image source: Alibaba Group.com
Alibaba has become a leader in open-weight artificial intelligence, rivaling startups DeepSeek and Moonshot. American Meta is one of the first players to open up the artificial intelligence market✴last week released a Glimmer model with 30 billion parameters that is small enough to run on a PC. Alibaba reminded the social media giant that competition in the market is fierce. The developers claim that the Qwen3.8-27B model has “excellent capabilities” in handling programming, research, and long-term agent tasks. Its stated answer quality is equivalent to a model ten times larger in size. Tests published by Alibaba confirm this – Qwen3.8-27B performs better than Meta✴ The Muse Glimmer 30B is just as good, if not better, than the closed version of Anthropic Opus 4.6 released in February.
Hays Resources reporters verified this claim. The original 32-bit floating point (FP32) model requires significant resources: at least 108 GB of video memory and approximately 120 GB of disk space for the KV cache. Model quantization up to FP16 allows you to run it on devices with 64 GB of video memory, while the more compact quantized version Q5_K_M requires 22 GB of memory – it can already run on flagship graphics cards AMD Radeon 7900 XTX and Nvidia GeForce RTX 3090 (TiForce), 4090 and 5090. It is also possible to use the option NVFP4 with weights, optimized for Nvidia’s Blackwell computing unit and runs with only 16 GB of memory.
As part of testing, Q8_0 and Qwen3.8-27B in the NVFP4 variant were launched in the Llama.cpp environment on a professional Nvidia RTX Pro 6000 Blackwell graphics card. These models are tasked with developing a complete REST API for an inventory management system, with consumer control and role-based authorization without asking any follow-up questions. The quantized version of NVFP4 consumes half the memory of the Q8_0 with no noticeable loss in accuracy or response quality.
Qwen3.8-27B writes high quality code – it compiles on the first try and supports all requested features. The quality of the project proved to be indistinguishable from that of the human Cloud. The Alibaba model took into account special aspects not mentioned in the task, discovered duplications in its own code during the work, and refactored them independently. There is some tendency to overthink simple problems, and the model becomes somewhat obsessed with details, resulting in wasted extra tokens and time. Although this behavior can be adjusted in settings.
AI models running on local resources have advantages over cloud-based models. There is no need to subscribe to personal data and transmit it to third-party providers. Anthropic and OpenAI’s closed models have no inherent limitations. Qwen also had some limitations, but shortly after the original version was released on Hugging Face, a derivative version appeared that largely eliminated these limitations.
If you find an error, select it with your mouse and press CTRL+ENTER.
