--- title: 'I ran 80B model on 16GB GPU - It''s surprisingly good! (Qwen 3 Coder Next Review)' source: 'https://youtube.com/watch?v=_UEvlmNx0cs' video_id: '_UEvlmNx0cs' date: 2026-07-28 duration_sec: 355 --- # I ran 80B model on 16GB GPU - It's surprisingly good! (Qwen 3 Coder Next Review) > Source: [I ran 80B model on 16GB GPU - It's surprisingly good! (Qwen 3 Coder Next Review)](https://youtube.com/watch?v=_UEvlmNx0cs) ## Summary This video tests running the 80-billion-parameter Qwen 3 Coder Next model on a 16GB RTX 5060 Ti GPU using a 3-bit quantized version. The model performs surprisingly well on moderate tasks but struggles with high complexity, compared against Gemini 3.1 Pro. ### Key Points - **Experiment Setup** [0:00] — Testing 80B model on 16GB GPU with 3-bit quantized version to fit VRAM. - **Model Introduction** [0:22] — Alibaba's Qwen 3 Coder Next has 80B parameters but only 3B active, enabling consumer GPU use. - **Configuration** [1:29] — 50K context length, half model offloaded to GPU, 1GB free VRAM, flash attention enabled. - **First Test: 3D Audio Visualizer** [2:00] — Model created a working 3D audio visualizer in one shot within 10 minutes. - **Second Test: Complex Web UI** [2:49] — Model handled most of a complex prompt but missed some details like blur layer and animations. - **Third Test: High Complexity** [3:30] — Model struggled with a very detailed concept, producing broken horizontal scrolling and layout issues. - **Final Test: Python Game** [4:47] — Model wrote a playable arcade spaceship shooter in 6 minutes, easily updated with sprites. ### Conclusion Running an 80B model on a 16GB GPU is surprisingly practical for moderate tasks, but high complexity still favors cloud models like Gemini Pro. ## Transcript In previous video, we tested the limit of 8 gigs GPU by running a 30 billion models on it. In this video, we're going to test it even further by running a 80 billion model on a 16 gigs card and then compare the result with Gemini 3.1 Pro. Will it works? Is it practical? We'll find out in this video. Couple weeks ago, Alibaba released a new coder model, Quen 3 Coder Next, and 80 billion parameters, yet score really well, almost on par with models much larger than its size. Quen 3 coder Next also has only 3 billion active parameters, which should allow us to run it on consumer graphics card with reasonably fast speed. And today, we're going to see if we can fit this model into a RTX 5060 Ti 16 gigs. So, here is my portable AIP PC. 32 gigs of RAM and 16 gigs of VRAM. It's natural that we have to use a quantized version in order to run it. Looking at its size, at first, the 4-bit quant version seems like it will fit. However, we'll be left without any spare system memory to use and very limited context, so it's a bit impractical. The 3-bit Imatrix version with 33 gigs in size should give us plenty of breathing room with almost the same intelligent with Q3KM version. And if you use Unslo quantized version, they even did some benchmark showing that their 3bit Imatrix quant performed better than non-unslaught Q4 and almost on par with the full precision model. For the setting, I'm using 50K context length and offload around half of the model to GPU, leaving 1 gigs of free VRAM. For the systems, I use flash attention but did not use KV cache quantization since the model itself was quantized enough as it is already. So I want to avoid any other potential quality degradation if possible. So the first test I'll ask it to create a 3D audio visualizer web app using 3JS particle to react to the sound using a blank web project as a base. Despite being an 80 billion model running on 16 gigs card, the token generation speed is pretty fast. The model completed the task in around 10 minutes and here is the result. It works beautifully. I'm really amazed I can get this result in one shot from local AI running on 16 gigs card. We can definitely polish it further, but for the first try, this is a very good result. So, the next test, we're going for a specific prompt. I found this good prompt example on X that he used it to test Gemini 3.1 Pro. So, let's use it to test with Quen 3 Next as well. The prompt again took only about 10 minutes and the result is pretty good for the first attempt. The model handled most of the prompt correctly with the exception of some. For example, it missed the blur layer for the glass morphism navbar. Some of the animation are not working and minor visual bug. However, if we compare the result with Gemini 3.1 Pro, even though Gemini handled the animation a lot better, Quen 3 coder next result doesn't really look that far apart. So moving on to the next test, we'll add more complexity by giving the model very detailed concept such as 3JS particles, horizontal scrolling, several animation, and different section layout. This time the prompt took around 30 minutes to complete and here is the result. This time we start to see the limit of the heavily quantized 80 billion model with complex concept prompt. The first section, navbar and menu look quite okay at first, but the other sections are quite bad, especially the horizontal scrolling, which is totally broken. Gemini handled this test much better. The final test. We're switching from web development to Python. The prompt is to write a simple arcade spaceship shooter game. The model took around 6 minutes to complete the task. And here is the result. The game is playable at the first attempt. And since I didn't provide the sprite image, the model used simple rectangle for the spaceship instead. So I provided them and told the model to update the code. Here is the final result. So running an 80 billion model on a 16 gig 5060 Ti isn't just a fun experiment, it's surprisingly practical. Even at 3bit quantized, Quentry Coder Next proved it can singleshot various tasks at moderate difficulty with impressive speed and accuracy. It still hits a wall on high complexity task which is where cloud model like Gemini Pro still hold the crown. But for local model on a consumer GPU, this is a massive leap forward. Let me know down in the comments what you want me to test next. And if you found this video helpful, like and subscribe for more. Thanks for watching. See you in the next one.