Tech2Wild @Tech2Wild

If you got 2 DGX Sparks and want to run Deepseek V4 Flash at 500k Context and be able to run it up to 8 Concurrency here is my recipe - github.com/tonyd2wild/dee…

Oh summarizing ? Say less then yes. I’m running 35B on my third 3090 for now and for conversations he wins he spits out response before I type the next question so summarizing I know it’s the same if not better. I need to start using him to summary YouTube video transcripts thanks for that idea

Made a decision today, and I am going 80% local. Moving ALL my agents to DSv4 Flash & Qwen 3.6 27B. Running my orchestrator, ONE Agent on Opus 4.8 to lead the charge. We will build automations & Skills that will help teach and shape the team. Workflows = Success. Wish Me Luck !

@jensenloke @FireworksAI_HQ @sgcodecampus No sir ! We can get you way better than that ! And with 3.6 Qwen you could run 27B and get damn near 4x the speed @aijoey he needs our help. Jensen join Joey group it’s a whole bunch of Spark users and a lot of people have just one Spark and they will get your right

@loktar00 Top of the week all the dust has settled. Qwen could drop 27B and it smoke DS4 and MM2.7 Kimi 2.5 and people would talk how crazy and so far ahead they are. Would push sales too of how good Qwen 3.7 is

@StefanMaier Hoping MM3 can fit on 2 sparks

I started noticing the local models that I was testing and using was capable of doing the stuff ChatGpt 5.5 was doing. And at times faster. Deepseek v4 Flash and Qwen 3.6 local speeds could accomplish the same things in the same if not shorter time AND FREE.

Randy Magruder @randy_magruder

7 hours ago

@Tech2Wild Just out of curiosity what made you make this decision

1 0 0 1K 0

@naturalSmartnes I’m new to Google and how they work because there whole tier shit is weird. Does this equate to usage if you have a subscription ? Or is this all API because that was my biggest issue with Google I never saw a usage bar.

3090. I have 3x 3090s, 2 are running 27B the 3rd is running 35B I think with a 5090 Frank you should be running 27B it’s smaller and more powerful you trade some speed for sure but with a 5090 you will will be good and with MTP it would be fast but the amount of extra power is worth it.

The same way AI companies are throwing compute and keep saying they want to slow down but competition doesn’t let them is the same way we should slow down lol. But think of it this way. Nvidia is about to drop a laptop that has a Spark in it. They are not going to tell consumers “don’t put valuable info in there”. I say that to say it seems weird now because we are the first to do it but this is the future. I remember my mom in the early 2000s not wanting to put her credit card on file for my Xbox account because she think it would get hacked. We gladly hand that stuff over now lol. Shit we tend to forget some of the subscriptions we are signed up to and still getting charged on 🤣

I bought a Mac Mini through the OPENCLAW Hype phase. I don’t regret and still keep mall my agents on it. I don’t see the negatives. I want one LOCAL computer that Harnesses all my agents and files dedicated to them. Am I missing something ?

Milo Smith @mil000

24 hours ago

THE GREAT OPENCLAW MAC MINI SELLOFF HAS STARTED!!

196 295 14K 1.2M 1K

@coltfeltes Yes someone mentioned and I haven’t tried it yet but I’ve thought about also doing this to video gaming. I love beating games by myself but I’m curious if I plugged Kai Into a game at a part where I’m stuck if I could have him help me get through a level.

@xmmooxx FP8

To have the feeling of my agent always being with me and on call. It’s one thing to type to them back and forth but what about an agent that can “see” and “hear” you and you can also “show” it what you’re doing. The more the agents knows the more it can help. And this all runs locally too. I’m going to release it soon try to get it out this week so people can have fun with it.

My agent Kai is watching the Nvidia GTC keynote right now via Eyes of Kai. He can see and hear everything as it comes in. The dope part is I can ask him questions about what is happening and he answers while still gathering the live speech to text in the background.

Ahh, ok so yes I saw those the selling part that made me go buy it was at the time “you can have you claw local and not on a VPS so all your files are there” and also it MAINLY was because ppl were scared it would take over ppls computers so running it on another rig isolated was better 🤣. Man I’ll be honest my OPENCLAW and Hermes has access to a lot more things now 🤣🤣

Thank you and yes I agree and yes I definitely make sure any model I run I want the ability to parellize. I have my opus agent cross check and sometime I even tell him to run subagents of himself to cross check. But I might need to also run other models as well other local models to cross check. Idk if there is a name for it but I do notice when I get multiple agents in a room and ask a question sometimes they tend to be easily persuaded by other agents too so that is my only fear. Keeping the checks in separate rooms might be better.

This is it, that’s the EXACT thought EC. Build with Opus and tell OPUS to make you a MD instructions that any agent on the team can read and know how to execute. Send that MD to the agent let them execute if they have an issue then have them prompt you what it is, send that back to opus and just go back and forth until it’s complete. Then have the local agent try agent once they get it right then you tell them to make that a skill and then you are now able to take that off opus plate. You and Opus are CEOs and Engineers the local are the workers.

@hishamfk This was me as well a few months ago it’s the right way to go and it will stretch your limits

@randy_magruder I quote tweeted this because it’s a good question.

Yes, I’ve learned a lot of the stuff is as people say is “theater” and I learned that “theater” looks cool but it over complicated and really slows things down. Trying to have 5 agents on one task doing an assembly line looks cool but one agent could do all of it in 1/4 the time. Definitely want to teach and do courses soon, I’m still soaking up game myself everyday !

I don’t NEED it I think at this point, but I feel like cloud will always be a step ahead and having that cloud model as the front runner it will allow and help me build better things and planning for the local ones. The cloud model is for building, planning, and strategy for what the local models are will run. Maybe in due time SOTA local will be here it’s close with Kimi and Minimax