Running LLMs on DGX Spark: DeepSeek V4 Flash Tops the Ranking
andersonbcdefg · x · 2026-08-03
A developer shared real-world benchmarks and rankings of serious LLMs running locally on an NVIDIA DGX Spark (128GB RAM).
- #1: DeepSeek V4 Flash 0731. Using 3-bit quantization, it runs at 16.5 tok/s. Although it's the slowest, it's the biggest brain (284b total / 13b active params) that fits into 128GB, making it a pro thinker worth the wait for hard problems.
- #2: Laguna S 2.1. Using NVFP4, it runs at 19.8 tok/s (up to 37 tok/s on code). It is the most trusted model for real agent work, capable of opening a todo list, tracking steps, and maintaining context over long hours without losing the thread.
However, the original thread was quote-tweeted by another user who joked that this complicated setup process is actually a fantastic advertisement to never buy a DGX Spark.
Related event: DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests(26 posts)→
More from Infra
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16