DevRico003

@UnslothAI @MiaAI_lab @0xSero Is there a possibility that it could run on 3–4 sparks?

You can now run Kimi K2.7 Code locally! 🌘 We shrank the 1T model to 325GB (-48%) via Dynamic 2-bit where important layers are upcasted. Run at >40 tok/s on 330GB RAM/VRAM setups. Run full precision on 610 GB. Guide: unsloth.ai/docs/models/ki… GGUF: huggingface.co/unsloth/Kimi-K…

I turned 100h+ of gut-health material into a Qwen SFT pipeline. But the important part is not the model. It is the loop: source chunks -> synthetic Q&A -> validation -> SFT -> eval failures -> repair rows -> retrain That is how domain models actually improve.

@Tech2Wild 🤯 this would be sick! I’d seriously consider buying two more DGX Sparks if this runs stable. Do you know if 3 DGX Sparks can be connected directly via QSFP, or would I need a 200GbE switch once I go beyond two?

@DevRico003 Minimax M3 on 3xSparks

Almost there

@aijoey That’s actually a smart way to enter physical AI. I’m also thinking about learning more robotics myself. That’s why I’ve been considering getting a Reachy Mini Lite from Hugging Face, connected to the DGX Spark, as a small local testbed for agents + physical workflows.

@DevRico003 I have no real project yet. I actually know very little about robotics. What I do know, is how to use agents. I'm breaking into physical ai with my agents as my guide. Looking into what seed studio has to see what I can buy and start deploying workflows.

quick context on cosmos 3. cosmos is nvidia's world model stack for physical ai. not just video generation. it is built to work across video, images, language, audio, and actions so robots and vehicles can understand scenes, simulate possible futures, and reason about what to do next. that is what this demo is trying to show: instruction -> world -> action one prompt becomes a physical scene, then a possible robot path. running this locally on the dgx spark is the fun part.

@ItakGol Even if this is fake, in times like these something like this could sadly become reality tomorrow. That’s exactly why we need to focus more on local LLMs stronger, lighter, and fully under our own control. Freedom will come from models we can actually run ourselves.

What the heck

@shawmakesmagic I completely agree. This will only make Chinese models and local LLMs stronger and make us more free and independent in the long run. I just hope the Chinese labs stick to their open-source strategy. Maybe this is exactly the wake-up call we needed.

China has the best model in the world right now (that you can get) This administration is a bunch of fucking idiots

@xikhar If true, this would be huge for European AI sovereignty. For enterprise and critical infrastructure like hospitals, US providers are often not an option. A strong EU model could be a real game changer and reshape the industry.

Le Chaton Fat

@aijoey @GZinMetaverse No, not yet, but after fine-tuning my own model with the DGX Spark, I’d definitely use the Spark to test the Locate Anything model too. That looks really interesting, especially if it can run locally and be useful for real robotics workflows.

@GZinMetaverse Very cool. Good job to you and your team. Are you playing with Cosmos 3 ?? x.com/aijoey/status/…

For the past 3 months, I've been heads down building and extending our multimodal data capture rig for Physical AI training. What you're seeing here is a single synchronized capture from our system — RGB, depth (stereo IR camera) feed; real-time 3D hand reconstruction, and a 23×20 tactile pressure grid from sensor gloves, all aligned and timestamped together. Getting here took a lot of unglamorous iteration. We reworked the pipeline to deliver clean, gap-free depth at scale after identifying early depth frame drops caused by data throughput bottlenecks in our capture box. We built per-channel calibration models that convert raw resistance signals from our flex sensor gloves into accurate, drift-corrected force values — solving the creep behavior that made a steady grip falsely appear to fade over time. Even though we’ve stayed quiet on socials, word of mouth spread — our rigs have already been deployed in data factories, manufacturing lines, and homes across the US, Europe, China, and Southeast Asia by our customers and ourselves. In the next few days, I'll be sharing more about how we're tackling rights-cleared data capture from the hardware level, and our closed-loop data-model flywheel — using model evaluation results to identify gaps and guide what data our customers need. More soon, stay tuned…

@iruletheworldmo This is exactly why local LLMs need to get stronger while staying small enough to fit on 1–2 DGX Spark-class machines. The goal is not just better models. It is freedom: local, private, uncensored, and not dependent on policy changes or shutdown risk.

on the bright side 5.6 is definitely coming next week and they make rather large donations to donald and take all of his calls so. there shall be no shutdowns.

@Gamingtronium Apple silicon

Hey devs, which processor is actually the best?

@ronisarkar_exe Codex

Developers, you have $20. Which one are you buying? – Claude – Codex – Cursor

@codewith55 WSL2

Just bought a new Windows laptop, What's the first app should i install?

@tunahorse21 I tested both the NVIDIA and ASUS versions. I ended up returning the NVIDIA Spark and buying the ASUS instead. If 1TB of storage is enough for you, the ASUS is a great choice for the price, especially compared to the more expensive NVIDIA option.

is there a big difference dgx spark vs asus ascent gx10? 5k vs 3.5k for what seems like the same hardware basically?

@alexocheema @nvidia 🤯🤯🤯🤯🤯

.@nvidia gave us all the hardware we need to make local AI awesome. 16 x DGX Spark 3 x RTX Spark 1 x DGX Station 24 x ConnectX-7 Cables 2 x High Speed Switches local dot AI

@mr_r0b0t @NVIDIAAI @spydenator Amazing!

@NVIDIAAI GB10 towers will never get old 😍 Gorgeous setup from our friend @spydenator!

The most important part of my SIBO / IBS fine-tuning project is not the training run. It is the eval repair loop. Every bad answer becomes: 1. a failure category 2. a better training row 3. a regression eval 4. another SFT run Fine-tuning is closer to debugging than magic.

I turned 100h+ of gut-health content into a Qwen 3.5 SFT pipeline. 200 videos. 1,891 chunks. 19,304 source claims. 33,705 train rows. 0 train/val overlap. Stack: - DGX Spark - Qwen 3.5 9B - DeepSeek synthetic Q&A - SFT + eval repair loop Goal: a source-grounded IBS & SIBO research assistant.