ChatGPT was the only model that insisted on creating raster images even though the prompt explicitly states generating an SVG output. Eventually I changed the prompt to "Please generate an SVG (not a raster image) of an octopus trying to solve a Rubik's Cube". The interesting is that it is also the only model that generated a real-time accumulated SVG so I could see the vector image taking shape as it was drawn.
These models can all be run on the average home computer on both GPU and CPU (although you'll have to be a bit patient if using CPU)
Ornith-1.0-9B was released on June 25, 2026, by DeepReinforce, and is post-trained on top of Gemma 4 and Qwen 3.5 foundation backbones Model GGUF size is 5.6GB
An open weights model Released by Google around July 2026. GGUF Size: 6.3GB.
Released March 2026 - GGUf size is 5.6GB
GGUF size: Qwen 2.5 Coder is considered one of the earlier most powerful coding models that could run on consumer hardware (Early means October 2024 - this is the pace we are at). I don't know if the abstract art supposed to be an octopus solving a Rubik's cube that it generated teaches us much about the improvement of coding abilities in newer models, but it is astonishing how much models on consumer machines have improved since its release.