BHCC Spark Lab

DEV and a nerd

[BHCC Spark Lab @BHCC2025]

@Tech2Wild Thanks for putting this together! Will check it out!

[BHCC Spark Lab @BHCC2025]

@ashxhart @NVIDIAAIInfra @vllm_project That would be awesome to see! Hope they grant access.

[BHCC Spark Lab @BHCC2025]

Two DGX Sparks, one model 🔗

Qwen3.8-27B (EXL3 4-bit) on TensorFold native, tensor-parallel across both Sparks:
62 → 92.6 tok/s on code (+49%), 128K ctx ⚡

Still testing 🧪
PR coming soon for a faster long-prompt kernel (128K prompt: 41 min → 3.7 min)

@ashxhart

[BHCC Spark Lab @BHCC2025]

@TeksEdge For 8K I was hoping for a memory bandwidth improvement.

[BHCC Spark Lab @BHCC2025]

@0xSero Now, those are some KLD number I like to see! I'll be checking that out!

[BHCC Spark Lab @BHCC2025]

@dumblebackdoor @ashxhart Glad to see the Arc support! It's a card that will gain popularity I think.

[BHCC Spark Lab @BHCC2025]

@Tech2Wild @theo

[BHCC Spark Lab @BHCC2025]

@GreymanITActual followed!

[BHCC Spark Lab @BHCC2025]

Nemotron 3.5 Super VL (120B MoE, 12B active, vision) running on a single DGX Spark. Official NVFP4 checkpoint, vLLM + speculative decoding.
First numbers: 43.6 tok/s decode, 2.5K tok/s prefill, 1M-token context fits.
Concurrency and tuning next, recipe to follow.

[BHCC Spark Lab @BHCC2025]

@GreymanITActual The future is bright!

[BHCC Spark Lab @BHCC2025]

I’m gonna give this a try. Claude was arguing with me about this that it wasn’t gonna make any difference until I showed him this post.

And then he said “oh that’s real data… and I should correct something I said earlier…”

That’s why I push back on these models - it’s what keeps them going forward!

[NVIDIA AI @NVIDIAAI]

Meet NVIDIA Nemotron 3.5 Super VL

Our new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents.

120B total parameters with 12B active 🧵

[BHCC Spark Lab @BHCC2025]

@ItsmeAjayKV @AntLingAGI KL divergance on some of the smaller quants looks good as well!

[BHCC Spark Lab @BHCC2025]

@ashxhart That's what I'm thinking as well - subsidized subscription plans won't last forever.

[BHCC Spark Lab @BHCC2025]

@WescheNex1q It's on of my "go to" models - looking forward to Qwen 4 27b

[BHCC Spark Lab @BHCC2025]

@digitalix Will be curious to see how this compares to the DGX spark - but I'm sure you are already on it!!

[BHCC Spark Lab @BHCC2025]

@DogukanUrker Glad to see LLM's getting more efficient!

[BHCC Spark Lab @BHCC2025]

Now testing: Naive-N0.5-Flash on two NVIDIA DGX Sparks 🔬

309B MoE, 15.5B active · TP2 · EXL3 3.5 bpw · speculative decoding · 200G RoCE between the boxes · pushing for the full 1M-token context

Benchmarking and tuning now. Recipe soon.

#DGXSpark #LocalLLM

[BHCC Spark Lab @BHCC2025]

Running Claude Code inside herdr, pasting screenshots into the prompt just doesn't work.

So I built a Claude Code mod: /shots

Newest screenshot is pre-selected, number keys pick more, Enter drops them into the prompt. Works with any screenshots folder.