Rejoice single DGX Spark owners! 💫 DeepSeek v4 Flash for single DGX Spark BEATS the 2x DGX Spark version on agentic workflows! 🤯🤯 Haven't done full coding tests but for agentic workflows it BEATS the FP8 version running on vLLM running on dual DGX Sparks. I fully expect it to be great at coding too! However, it's much slower, as expected: Aggregate tok/s (1-2-4-6-8-12) 26.7 tok/s → 32.9 → 46.5 → 54.1 → 58.5 → 58.5 tok/s Per-stream progression: 26.7 → 16.5 → 12.1 → 9.5 → 7.7 → 5.2 tok/s That means up to 59 tok/s across 12 concurrent sessions. This is the best model to run on a single DGX Spark! All credits go to @bleysg! I have published a simple start/stop recipe: I still recommend running DeepSeek v4 Flash on 2x DGX Sparks because the speed difference about 3x. Additional details below 👇

DeepSeek v4 Flash for 2x @NVIDIAAI DGX Sparks has been updated✨ - Single stream recorded ~72 tok/s decode - 6 concurrent up to ~137 tok/s aggregated It's still the best model to run on 2 DGX Sparks. →

You can now run GLM-5.2 on 3x @NVIDIAAI DGX Sparks at home 🤯 • 248k context • 15-20 tok/s (content-dependent) • MTP for max context / DSpark optional • NVFP4+AQLM hybrid Still a work in progress — will keep improving over time. →

I think @PrismML is onto something. Bonsai 27B is already exceptional at tool calling. If their next release is properly trained for coding, it could be a genuine game changer — especially for users on 12-16GB VRAM GPUs and laptops. I'm rooting for them.

This model can fit in a 12-16gb VRAM consumer GPUs and laptops and it's almost as good as Nvidia's Qwen3.6-27b NVFP4 for agentic workflows, which is insane 🤯 @PrismML's Bonsai-27B 2-bit is performing exceptionally well! Coding tests next. Full eval results link:

Here's one way to get the most from Grok 4.5 ✨ Switching effort to "low" gives strong results on most tasks, near-zero quality loss, big usage savings — plus it's the fastest.