Hi, I’m currently getting to grips with a pair of Sparks. Maybe some of you have already gone through the same thing, so I’d like to ask for some tips regarding VScode and VSPro. While VLLM models work fine in VScode or through LiteLM, VSPro keeps timeouting (I’m using a proxy, I’ll see). The cursor that came with the package doesn’t work at all - does it require Pro to connect the host with the models? Another question is, are there any proven model combinations for development? Or do I have to find them myself? Any advice is welcome.Thanks :)
I have a hard time understanding what is a problem or has to be done.. VSpro is ? Visual Studio 2022/2026? You want to run it on Spark itself? What Cursor has to do with it? Sorry I cannot help, just puzzled what is going on
hi just wanted to integrate copilot at the both and somethow it works via proxy and litellm. It is just so simple so i am wondering about any ‘magic’ setup - yes its pro 2026. Cursor i just wanted to try this agentic coding there as i heard its quite nice but at the free version seems that this will not work.
I use VS 2026 (and 2022 before). There is NO WAY to make copilot work with local model - either github copilot of couple of affiliated providers. Design choice. I wonder why you need that. It was a choice like 1-2 years ago when AI was used for autocomplete or block generation. Now use Github Copilot app (can be perfectly integrated with local or any inference with openai api) or better still ditch the ms/github bloat and move to OpenCode/Pi/Jetbrains etc
My workflow for C#/DotNet is I keep VS 2026 open and OpenCode for agentic work open, same repo/directory/project. Agent works in OpenCode, I look into VS 2026. They syncronize perfectly. Python - OpenCode (or Hermes better) + PyCharm for my eyes only (lol)
ah ok so thats what i was looking for :) opencode and simulatenously working with the vs2026 :) I just used to claude and vs copilot - for my kind of work / big sensitive system it was working really well but now i needed to make the step forward to let model look at whole repos etc. So thank U for the hint :) Any model suggestions? All i know for now is that i need some cooling chamber for those two bastards :)
Right now the best model for 2 sparks that can code well and handle 1M context is deepseek v4 flash. Nothing comes close. If you prefer multimodal and can work with 256-300k tokens - Qwen 3.5 122b will be excellent choice too, very fast on 2 sparks. 397b is hard to fit, nearly OOM, tiny cache, not that much better, no reason to struggle. Minimax M2.7 is fine if 200k session limit is not bothering you.
Thank U very much. All stuff is going on now. :)