Ollama 0.15.5 new release

New models are now available, such as qwen3-coder-next.

This model writes code very well, but there is one problem. For Ollama, there are only models with quantization q4_K_M with a model size of 52GB, q8_0 with a size of 85GB, or a paid cloud launch. The model itself is trained on 80 billion parameters, most of which are code.

Yes, in the minimum configuration, at least 80GB of video memory is required for local launch of this model if high speed is needed. Of course, you can run this model on CPU with at least 128GB of DDR5 RAM, it will work, but very slowly.

Interestingly, such articles always attract haters; I think it’s because not everyone can assemble a server to run such models. But the problem is that when I offer such people access to my server, which already has these models loaded and has enough power for their inference, they start downvoting the articles and writing ridiculous comments about how Ollama is complete garbage.

If you are not a hater and not a super guru in neural networks, I would be happy to give you access to my AI server where you can try out 10 different models.

Access to the server is free, but unfortunately not permanent. When I finish setting up my RAG, I will turn off this server. Although there is a slim possibility that I will bring another one like it from China.

The server address is http://demonryb.ru:8080/

Access to the models is only possible after I confirm your registration.

Just to add, if I approved your registration BUT you try to overload the server with stupid requests, I will be forced to disconnect you.

And here is a brief specification of my server:

Processor: Core i9 14900KF

RAM: DDR5 192GB.

GPU 2 x RTX4090D48G (96GB VRAM GDDR6X), performance according to huggingface standards 166 TFLOPS

This may not be the most successful configuration, I’m ready to discuss it, but if you don’t even have such a configuration for your PC, maybe you shouldn’t shout and downvote me?