We are glad to announce the immediate availability of Gemma-3, a family of vision-capable models released by Google in March, and QwQ-32B, a medium-sized reasoning model from Alibaba. With the arrival of the former, SCRP-Chat now offers a full range of vision models. We will therefore retire our existing range of text-only Llama 3 models one week from today (April 21th 2025).

In addition, models on SCRP-Chat can now search the web and generate images:

  • Web search is disabled by default. You can turn it on by clicking on the ‘+’ button at the bottom of the prompt textbox and switch on ‘Web Search’.
  • Image generation is disabled by default. You can turn it on by clicking on the ‘Image’ button at the bottom of the prompt textbox.

Please refer to the SCRP-Chat documentation for usage instructions.

Gemma-3 27B/4B

Gemma-3 is a family of vision-capable models from Google.

  • The 27B version is the largest and the most capable in the family. In our internal test, it is competitive with Pixtral Large in many areas. The latter is stronger in mathematics, but reasoning models are even more capable on that front.

  • The 4B version is the smallest vision-capable model in the family. In our internal test, it runs much faster than its larger sibling, but its ability to understand tables is noticeable weaker. 

Generation speed:

  • 27B: up to 44 token/s.
  • 4B: up to 76 token/s.

QwQ-32B

QwQ-32B is a medium-sized reasoning model from Alibaba. It significantly outperform DeepSeek R1 Llama 70B on benchmarks, but in our internal test, it also takes significantly more time than the latter in solving mathematics questions.

Generation speed: up to 40 token/s.