I'm amazed - local 16G runs Qwen38-27B-UD-IQ3_XXS

Test: Qwen3.8 Flash Next vs Qwen3.8 27B! Qwen3.8 Flash Next Effect

Local deployment of qwen3.8-27b-gsq is amazing. I recommend it to AI enthusiasts.

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.

Claimed to be a quantized IQ3 model benchmarked against full-blooded BF16, Qwen3.8-27B-GSQ-RCO-GGUF

Qwen3.8 27B: The gap between different KV cache quantizations, trying to fill in the last piece (of the puzzle) of the evaluation.

I was fooled again by the Bonsai-2-27B model. I thought it was based on Qwen3.8-27B, with a context window of up to 262K, could see images, and call tools. The officials claimed it retained 98.2% of the original intelligence.

The ultimate thousand-question deployment plan for ordinary graphics cards = kvmem + bonsai27b

amd yes! Using the 7900xtx to push qwen3.8-27b to 97t/s! 256K context with vision!

DeepSeek V4.1 generation speed

Qwen3.8-27b (Q5_K_XL/Q4_K_M) vs v4flash vs v4Pro

Local model Qwen3.8 (choose 27B or flash-next-strata)

Please choose your quantization! Qwen3.8 27B Q8XL Q6XL Q5XL Q4XL IQ4XL YMQM YMQCFGM XLGSQRCO-IQ3S Test them one by one!

Bravo! Qwen3.8 27b, this episode is amazing!

Windows11 native VLLM deployment of Qwen3.6-27B at 100 Tokens per second

Qwen 27B local deployment test: 8897 Tok/s, truly unbeatable?

Running the qwen-3.8-27B-IQ4 model on a 16GB VRAM graphics card, with MTP and vision enabled, reaching a speed of 50 tokens per second

If you only have a 16GB graphics card and want to barely run the Qwen3.8-27B quantized model locally, you can try the quantized version from ISTA-DASLab's GSQ-RCO.

The era of AI is coming, ubuntu26 will surely rise！！！！！！！！

Qwen3.8-27b + DSH feels similar to or even better than Deepseek-v4-flash + DSH