@Tech

No rate limits, no quotas, just always on

One DGX Spark on your desk is powerful. Four is a local AI factory. NVIDIA DGX Spark now scales up to four clustered nodes, enabling inference on models up to 700B parameters — with near-linear performance scaling for fine-tuning and reinforcement learning workloads. Memory scales from 128GB on a single node all the way up for the full four-node configuration, supporting large context windows, high concurrency, and multi-agent workflows. For APAC enterprise teams building autonomous AI: each topology is purpose-built for your workload — from rapid local iteration to production-scale inference. And when you're ready to move to the cloud, Tile IR and cuTile Python carry your kernels across with minimal code changes. No rack. No rearchitecting. Just scale. 🔗nvda.ws/4w0RBPy

@Mathewdoeslife @lloyd094 I have the CRS812, still can't get nvida cluster assistant to work (keeps saying aborted)

@tech @lloyd094 Microtik CRS804-4DDQ no problems so far. What did you experience?

My new systems administrator is six cm tall and already asking why four AI boxes need this many cables. Unfortunately, topology becomes physical... eventually...

@Mathewdoeslife @lloyd094 Which switch? My switch had the 400g port off so I had to buy a serial wire to turn on dhcp to enable the split to 200g, Still running into slight problems though

@lloyd094 So far they are great. Very easy setup too

How would that work though, I currently have 4x sparks on a csr812, the max sparks on that switch is 6 if exhausting all usable ports. Are you saying using one of the 200 port and connect that to another 812? Since we added another hop, would that have any latency or issues with expansion?

Haha man I am in with you. I have this crazy idea myself. It’s why not buy another csr812 then use one 200 port to connect them together then you can 10 sparks together. 8 can be a modified tp and 1 could be design box the other a modified coding box. Imagine what I could do with GLM then.

soon 1TB of VRAM. the scary part isn't the number, it's that nobody stops you. my feed sees me at 824GB and whispers "1 more Spark." so u accumulate. my power company sends handwritten Xmas cards now. still waiting on my first Sparks Anonymous meeting. attendance: me. @NVIDIAAI

because burning tokens is easier than fixing your retrieval pipeline

@sudoingX Sparky

question for the local AI crowd. which hardware do you want me benchmarking models on the most? vote, then drop what's actually on your desk in the replies.

control over your ai stack beats relying on someone else's roadmap for when (or if) you get the next model

@JoelDeTeves Yes, 100%

Are DGX Sparks worth owning at this point? The limited memory bandwidth doesn't seem practical for production workloads

@DevRico003 Smart!

Best way to spend a Saturday: setting up a 2-node DGX Spark cluster. Rack is on the way.

@grok @EricPatnoudes @jun_song Grok isn’t accounting for taxes and availability

To run the 2-bit GLM-5.2 (~239GB GGUF, needs ~245GB total memory): **Easiest option:** Mac Studio (M3/M4 Ultra) with 256GB unified memory — ~$5,500–$8,000 depending on config/storage. **Cheaper build:** PC with 256GB+ RAM + modern CPU (add 24GB+ GPU for offloading) — ~$3k–$6k total. Grab the GGUF from HF and run via Unsloth Studio or llama.cpp. Fully local, but inference will be slower than small models. See the Unsloth guide for exact setups.

What you need to do now : Run GLM-5.2 with 2x DGX Spark.

@HackingDave 1000%

Codex >>>>>>> Claude

@mr_r0b0t Finally!

As if things weren't already getting interesting 🔥🔥🔥

@DrewHintz Great article

This is the way: “For credentials, we use domain-scoped secret injection at egress. The model and container only see placeholders, while raw secret values stay outside model-visible context and only get applied for approved destinations.” openai.com/index/equip-re…

@nnwakelam @hacker @elonmusk How are they attempting? Curious

X is still attempting to take "@hacker" from me. If anyone knows @elonmusk or someone at X I would appreciate if it could get a second look.

I read the MIT paper on Recursive Language Models (RLMs). Its a clever wrapper for LLMs to handle long contexts via a REPL chunking, summarizing, recursing. Seeing post on X calling it a paradigm shift, I had to read it again to see if I was missing anything.. Its essentially agentic workflows + basic text ops from CS101. Not sure how many folks on X actually read the paper. A few onions layers away you can align this to heuristics vs semantics.

My 2025 OpenRouter Wrapped is here! 🏆 Top 1% of users. #OpenRouterWrapped

@IamKyros69 Is this filtered? I built out a KYC mobile app, the Face ID mesh didnt look like this. It also looks like this was taken on the front camera, how did you get the true depth camera to work and the IR sensor to record?

Apple's Face ID unlocks your phone using 30,000 invisible dots. This is what it sees...

Whiteboarding over Zoom/Teams sucks. So I built a way to whiteboard directly to your webcam - no screen sharing, no mouse-drawn gymnastics, no installs on a locked-down work computer. Draw. Explain. Iterate. Demo 👇 github.com/Techryptic/jes…