GLSL Shader generation with the mini preview model
I did some extensive testing today with a GLSL shader request that I've run through quite a few models. Right now, if I describe a scene, Qwen3.8 Flash-Next · GSQ-RCO IQ3_S-xhigh does a very good job and a pretty fast turn around. (about same speed as Astra-Max with about 85% the quality, subjective, but I've run many prompts to compare).
Today I tried to do this same test with Raon-1 mini · Q4_K_M, Q6_K and Q8_0 with native on/off and I am not getting anywhere near recognizable for the prompts. More, a lot of the time the code is broken and has to be repaired by additional calls.
Anyone have any idea of what I might have wrong with using this? Looking at the benchmarks I was expecting something a bit better. But, given this is a mini model, maybe I will just need to wait for the full release. Ideas?
Hi grimepoch, thanks for sharing your experience with Raon-1 mini preview. We hadn’t tested GLSL shader generation specifically, so it’s helpful to hear about your results.
Could you also try the same prompts on our free official playground?
https://raon-mini.krafton.ai/en
If they work well there but not with GGUF, let me know. I’ll check our GGUF implementation and see what needs fixing.
And if you don’t mind, could you share your test prompts and setup? We’d like to take a closer look and consider adding examples like these to the training data for the full release or future models. You can post them here or email me at hyogon.ryu@krafton.com.
We’ve also opened an official Discord server, so feel free to join us there. https://discord.gg/RMRjbBPJWz
Hello ugonfor! I did look at the discord, but I didn't see any English channels so I wasn't sure where to post.
This is the prompt: "Make a GLSL shader. Follow the shader format at glsl.erogenous-tones.com. Create an underwater observatory grown from luminous coral around a ruined astronomical instrument. Make a composed 3D scene with foreground, middle distance and a hazy background: curved stone ribs, a detailed focal instrument and smaller organic growth at multiple scales. Use procedural geometry with real occlusion, surface normals, distinct stone/metal/bioluminescent materials, soft lighting, water caustics and depth-dependent absorption or fog. Animate drifting particles, caustics and a slow instrument mechanism continuously without moving the whole scene as one texture. Keep silhouette and focal point readable. Controls: water clarity, growth/structural detail, and bioluminescence intensity. Keep a useful, attractive scene at all control endpoints."
There might be some other stuff around this as I have astra generating these runs.
Here is an example of the output from the Qwen3.8-27B Xhigh thinking · 256K context · Q4_K_M - 65 minutes
This was done with Qwen3.8 Flash-Next · GSQ-RCO IQ3_S Xhigh - 11.7 minutes
Which is my current favorite.
For your model, Raon-1 mini · Q4_K_M - 32K output
I of course ran lots of the settings but almost none of the results had any relevance to the prompt.
For the console model free to use on the site, I tried that this evening. First pass, the shader would not compile. So I asked it to repair it. Second pass, still failed. Third pass it decided to look at the website. after about ten passes I just stopped trying to get it to fix the shader.
Thank you for sharing your detailed experience, including example prompts and outputs!
Our current model wasn't designed for vision tasks, so it may not perform well on them. We'll take these use cases and examples into account for our future model. We’re considering vision support for future models, so please stay tuned and give our next model a try!
Thanks again for sharing!
Understood! I didn't even consider the vision side of it with this. I will keep my eye on your future models for sure.