@Tech2Wild I have one but the problem is I also have a rtx 6000 so one card spoils me with speed and the spark just makes me want to buy more 6000s so that I can get that speed with bigger models
@jun_song You are right. The only “consumer” card that is worth it for local inference outside of maybe a 3090 is the rtx6000. But as someone that has a rtx6000 + a spark. I dread if I have to put a model in my spark because it’s so much slower than the 6000 😭
@analogalok@ArtificialAnlys Are you running it locally? because 27b xhigh is going to think for 10 hours before it spits anything out. for agentic tasks I have to set thinking to off if I want any semblance of speed, as well as very clear cut no room for interpretation instructions.
@Hikari_07_jp I'm running Qwen3.8-27B-Uncensored-FP8. It works great just thinks way too much. Super good with a different model guiding it and giving it specific instructions.
@StocksOnSpaces I don’t know if anyone reads this but one barrier to entry for AMD and even Apple silicon if they open it up. If they aren’t backwards compatible with NVIDIA GPUs they have to be significantly better for people switch over.