Qwen3-Next

Take math and coding for example:

- in math, if they can solve a problem, or a class of problems, they'll solve it. If you use a "thinking" model + maj@x, you'll get strong results. But if you try for example to have the model consider a particular way or method of exploring a problem, it'll default to "solving" mode. It's near impossible to have it do something else with a math problem, other than solving it. Say "explore this part, in this way, using this method". Can't do it. It'll maybe play a bit, but then enter "solving" mdoe and continue to solve it as it was trained.

In practice, this means that "massive parallel" test time compute becomes harder to do with these models, because you can't "guide" them towards certain aspects of a problem. They are extremely "stubborn".

- in coding it's even more obvious. Ask them to produce any 0shot often tested and often shown things (spa, game, visualisation, etc) - and they do it. Convincingly.

But ask them to look at a piece of code and extract meaning, and they fail. Or ask them to reverse an implementation. Figure out what a function does and reverse its use, or make it do something else, and they fail.

This model can be run completely offline, yes. You'll need anywhere from 60-200 gb of RAM (either VRAM for high speeds, or a combination of VRAM and RAM, or just CPU+RAM). The active params are really low (3B) so it'll likely run fine even on CPU. Should get 10-15+t/s even on old DDR4 systems. Offload some experts to a GPU (can be as low as 8-16gb) and you'll see greater speeds.

This has nothing to do with nano banana, or image generation. For that you want the qwen image edit[1] models.

1 - https://huggingface.co/Qwen/Qwen-Image-Edit

It sounds like some people.

What kind of benefit does Multi-Token Prediction bring to the inference side? Is it only relevant in pretraining efficiency?

llm -m qwen3-next-80b-a3b-thinking "An ASCII of spongebob"

Here's a classic ASCII art representation of SpongeBob SquarePants:

     .------.
    /  o o  \
   |         |
   |  \___/  |
    \_______/


  llm -m chutes/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 \
  "An ASCII of spongebob"

  Here's an ASCII art of SpongeBob SquarePants:

      .--..--..--..--..--..--.
    .' \  (`._   (_)     _   \
  .'    |  '._)         (_)  |
  \ _.')\      .----..--.'  /
  |(_.'  |    /    .-\-.  \---.
  \     0|    |   ( O| O) |    |
   |  _  |  .--.____.'._.-.  |
   \ (_) | o         -` .-`  |
    |    \   |`-._ _ _ _ _\ /
    \    |   |  `. |_||_|   |
     | o |    \_       \    |     -.   .-.
     |.-.  \    `--..-'   O |     `.`-' .'
   _.'  .' |     `-.-'      /-.__   ' .-'
  .' `-.` '.|='=.='=.='=.='=|._/_ `-'.'
  `-._  `.  |________/\_____|    `-.'
     .'   ).| '=' '='\/ '=' |
     `._.`  '---------------'
             //___\   //___\
               ||       ||
               ||_.-.   ||_.-.
              (_.--__) (_.--__)

Meta: I generated a few dozen spongebobs last night on the same model and NONE where as good as this. Most started well but collapsed into decoherence at the end - missing the legs off. Then this morning the very same prompt to the same model API produced a perfect bob on the first attempt. Can utilization affect response quality, if all else remains constant? Or was it just random luck?

Edit: Ok, the very next attempt, a few minutes later, failed, so I guess it is just random, and you have about a 1 in 10 chance of getting a perfect spongebob from qwen3-coder, and ~0 chance with qwen3-next.