hi @altamiro.aj
I’m currently using several models like QWEN 3.8 Flash Next 120B and GLM 5.3 320B. Really appreciate for mentioning Physalia. Just added to my list to try in my experiment.
hi @altamiro.aj
I’m currently using several models like QWEN 3.8 Flash Next 120B and GLM 5.3 320B. Really appreciate for mentioning Physalia. Just added to my list to try in my experiment.
Just out of curiosity, whats the VRAM/RAM usage on this kind of models?
Maybe we could split the thread or I should create a new one, sorry!
But I keep thinking about the example @Gijs shared.
Have any of you guys have experience with AI (claude or chatgpt) in creating actual GH definitions with no python scripts?
Because I think AI is currently really good at python but most suck with Grasshopper…. ![]()
Also whats the difference between RHINO MCP and RAVEN PLUGIN? Is either better to create grasshopper scripts?
I’ve used DeepSeek with Physalia for this, it worked fined
I tried using Claude without Physalia but it create some monstruocities that were awfull.
Raven has some amazing capacibilities, it can create the GH definition as well as a interface to use on Rhino without having to mess with GH Canvas. Raven is having a double free tokens until the 31st, its an amazing moment to test it.
My laptop setup of multiple GPUs carries around 100GB of VRAM (alongside my Radeon iGPU Smart Memory at 57GB) and 80GB RAM left. With GLM, it took almost both memories and left enough to do just windows and emails ![]()
But QWEN left more so I could still do my other stuff while it carried Rhino’s modeling task.
I’m now setting the Radeon to carry 76GB or VRAM through Smart Memory while securing 8GB to do windows graphics.
Interesting! Thanks for sharing your experience so far!
I wonder how is Raven different than Rhino MCP if at the end both are using the model you select say Astra from ChatGPT… technically both are as powerful and capable? What is that Raven adds in making it worth the extra subscription? Do you actually need to pay Raven and ChatGPT or does it coke included?
As far as my understanding, Raven is the complete solution, you do not have to have both subscriptions at once.
I don’t think it will have much difference for now appart from this hability to create customized UIs for use and all, but in the long run, I am sure Raven will evolve to be a more specialized and powerfull tool compared to general models.
Even more that they are trying to limit the models for cybersecurity, there will be consequences for coding and programing with CAD software for sure.
In the last 2 days, I’d been comparing the result between Rhino 8 and Rhino WIP AI generated objects.
It seemed that Rhino WIP did the modeling better than Rhino 8. Some details were captured by Rhino WIP but skipped by Rhino 8.
My preliminary conclusion is, with local AI, it may be carrying similar situation to “a person who was asked to look at an object once, and being asked to sketch, draw or even modeling it without giving that person chances to look at it again to keep on adjusting details and proportions of the object he/she saw.”
For reverse engineering with layout drawings, the local AI is sufficient at certain levels. This, gives me confidence that in the near future, there will be many more genius individuals or organizations that may release open weights, far beyond what we’ve seen today. And this could bring a new horizon for us, Rhino users, to benefit from.
There are more things I need to experiment with my current setup. The long couple of weeks brought me joy playing with Rhino on top of a new ground with local AI.
Keep Rhino-ing friends!!! ![]()
Thanks for all these investigations @bimolicense, they’re very interesting to follow along with.
I want to do the exact oposite
I do the 3D modelling and I want AI to act as my intern and work on the workshop drawings.
So far, not much success with this part ![]()
@Gijs When clade built this? Is it live on the canvas? or did it deliver a GH file you then opened?
That was built live.
Really appreciate the kind words @CallumSykes. Frankly speaking, I’ve explored and worked with various 3D software since my college days back in the early '90s, and first encountered Rhino later on right after Frank Gehry relied on Rhino was announced. By 2010, I had completely fallen in love with it, and that passion has only grown stronger.
I’m deeply grateful to be part of this community, and particularly thankful to the seasoned users in this forum whom I continue to learn from endlessly. The spirit of sharing and mutual support here is a powerful magnet that keeps me coming back.
Your words only anchor me further, firmly welded to this community and to Rhino ![]()
Updates: On several random cases, MCP still launched new Rhino window and skipped currently opened Rhino WIP. In this case, once the task was completed, the new Rhino windows was immediately closed.
Any particular known way to prevent this other than signing port 10501?
I started to experiment with scanned image to layout drawing with my setup.
At certain level, it transformed the image into 2D layout drawing. It’s a mixed bag of accuracy. I will continue to experiment with more trials and details, and so far, it’s another interesting thing to move forward ![]()
Few things that made me surprised were layers were created accordingly, objects were placed to each related layers, and dimensions were known to its own categories (radial dim, linear dim).
Today’s progress.
Showed the ecosystem to a friend. An architect. His face was pretty much like crumpled paper when seeing not just Rhino window, but terminal, dashboard on browser and VS Code. For him, VS Code is pretty daunting ![]()
So, I read the list on Rhino MCP Platform, trying to identify which one supports local AI.
After some trials, OpenCode seems to work best with local AI models and even LM Studio. However, integrating it to Strata with n-gram, was head scratching. After couple of hours, connection between Strata-OpenCode-Rhino was established
Didn’t think this could happen.
Made some time measurements, seemed like OpenCode ran the process for Rhino faster than VS Code.
More to come.
UPDATE: trying to add OpenCode as agent in Rhino AI panel, but still no luck
I guess I’m continuing with OpenCode GUI at this stage.
Some trials confirmed OpenCode interface is faster than VS Code when interacting with Rhino, whether it’s carrying none, low, medium or high reasoning.
On several attempts in modeling the bracket, with some changes in prompt, the results is always the same. However, reasoning setting on default, medium or high is required to get the shape and accuracy done correctly.
I may try to let it observe 3d object and then ask it to do layout drawing later ![]()
Oh, one more thing. While the AI is in the middle of processing the instruction, a new instruction could immediately be inserted to carry more adjustments or add another instruction without the need to wait for the ongoing to finish.
UPDATE: This was done through conversations in between accomplishments, not just single prompt.
During my experiment in providing components to complete the rendering process, like generating textures automatically without cloud AI, I found a very interesting thing.
It shocked me deeply, knowing that it could now look at what inside the GPU and its VRAM. TLDR: it completed all things that I need to be utilized in Rhino, by itself as well as balancing the usage and loads at the same time
just by using local AI.
It gave me a long deep goosebumps.