V100 running Qwen3.8 27B Qianwen nvfp4 with speed over 50 tokens, about half the speed of Pro 6000d. In the current environment of soaring graphics card prices, it is probably the most cost-effective AI operation graphics card.
99% of locally deployed AI are not as good as Doubao
I'm amazed - local 16G runs Qwen38-27B-UD-IQ3_XXS
Qwen3.8-27B disrecommend, meaningless for average users.
Ridiculous! Running Qwen3.8-27B on 16GB VRAM, with 126K context, 40 tokens per second, and even bf16 quality KV cache precision
Scared to tears! Qwen 3.8 27B's ridiculous performance; can you believe this is only a 27B model? | Small Wave Mom's Radar Chart
Qwen3.8-27B is officially open source! Intel running Lowest 8GB VRAM required is enough to run it. Can it rival the top closed source models? Localization test recap | Zero Degree Explanation
Running Qwen3.8-27B on a single 16GB graphics card, local AI is really good!
Using only 5GB VRAM! Deploying Qwen3.6-35B-A3B + Visual Model Locally on a Laptop with RTX 3060 6GB
Ordinary people can also play Local AI Deployment! Qianwen 3.8 27B + DeepSeek Harness installation tutorial, follow along and you can run it
【Save Money Tips】A 500-yuan NVIDIA V100 16G can run vllm: qwen3.8 27b, with speeds up to 200 tokens/s
v100-32g can only run at around 25 speed, qwen3.8-27b
Someone still running qwen3.8 27b with q4 quantization complaining about high hallucinations? A minute to explain what exactly is the difference between q4 q6 fp8 bf16 nvfp4!?
Running a 27B large model on 16GB VRAM? Qwen3.8-27B local test: 40T/s, this time it's truly legendary!
Please choose your quantization! Qwen3.8 27B Q8XL Q6XL Q5XL Q4XL IQ4XL YMQM YMQCFGM XLGSQRCO-IQ3S Test them one by one!
Local Deployment Qwen3.8 27B: The Strictest Pelican Test Is Here
[Chinese Dub] Qwen3.8 9B Model In-Depth Test: Small yet Powerful, Close to Large Model Performance
650 yuan mining GPU runs Qwen 3.8 27B model - speed close to 30 tokens/s
27B Only 6.6GB: Qwen3.6-27B Ternary Edition, Ternary-Bonsai-27B
5060ti 16G boldly running 35B model, 80 tokens/s
