Tag
An observation of Ling-3.0-flash (124B) on one DGX Spark generating 15,128 tokens in a single response with stable decode throughput around 35.6 tok/s, highlighting long-context decoding performance.
An independent benchmark by sudoingX shows the Ling-3.0-flash model runs at 38.7 tok/s on a single DGX Spark with official INT4 quantization, 2.4x faster than DeepSeek V4 Flash on the same hardware, after a correction clarifying the quants do work.
User tests Meta's Muse Glimmer 30B on a 2× DGX Spark cluster, extending context from 131K to 1M tokens with YaRN and confirming passing retrieval at 832K tokens. Reports ~3× speedup from DFlash speculative decoding and shares full config.
This post describes two configuration flags that increase the official Ling-3.0-flash INT4 inference speed from 20.8 to 38.7 tok/s on a single DGX Spark, while warning about the need for a specific vLLM fork and noting tradeoffs with long-context performance.
Ahmad Osman argues that dense models like Qwen 27B perform poorly on unified-memory systems such as NVIDIA DGX Spark, suggesting MoE models are a better fit; he claims discrete GPUs like the RTX PRO 6000 deliver far better performance for agentic workloads.
User seeks community advice on reducing VRAM usage and freeing OS RAM when serving DeepSeek-V4-Flash-0731 on two DGX Spark machines with vLLM, sharing detailed configuration and memory measurements.
Ray Fernando expresses excitement about Local Studio, an open-source interface for running AI agents on two connected DGX Sparks, similar to Codex.
A practical guide to selecting MikroTik switches (CRS504, CRS804, CRS812) for building DGX Spark clusters, including pricing and real user examples of 8x DGX Spark setups.
The author ran same-conditions tests of MiniMax H3 and LTX 2.3 Eros video generation models on DGX Spark. Results show H3 is stronger in narrative and instruction following and includes audio, while LTX is about 1.76x faster.
A 14.7-hour soak test of DeepSeek V4 Flash on an NVIDIA DGX Spark with 971 requests shows zero crashes or errors, but throughput declined 28% due to thermal throttling, while TTFT and speculative acceptance remained stable.
A user shares that they replicated running DeepSeek-V4-Flash-0731 on a DGX Spark using antirez's DwarfStar-4 setup, confirming impressive performance on a single device.
This open-source project enables Nix and NixOS on the NVIDIA DGX Spark, providing USB boot images and a NixOS module for configuring the hardware.
A user is setting up a 16-node DGX Spark cluster to run frontier open models locally, discussing networking tradeoffs between 200 and 100 Gbit/s.
A detailed strategy using VRAM as disk cache with unified memory and llama.cpp settings achieves 340 pp/s and 9.6 tg/s for Kimi K2.7 on a single DGX Spark.
The author reviews the NVIDIA DGX Spark as a capable general-purpose computer, praising its performance, power efficiency, and ability to run games like Crysis, despite being primarily an AI development device.
MiaAI Lab released an upgraded recipe for serving DeepSeek V4 Flash on two DGX Spark nodes using vLLM with DSpark speculative decoding and NVFP4 KV-cache, achieving up to 153.7 tokens per second with six concurrent sessions.
User tests PrismML's new Bonsai 27B model on Nvidia DGX Spark, reporting benchmark speeds and issues with llama.cpp build, while PrismML announces the model as the first 27B-class model to run on a phone.
Describes deployment and benchmarking of the quantized GLM-5.2-Int4-Int8Mix model on an 8-node DGX Spark (GB10) cluster using a custom vLLM fork, achieving ~1,200 t/s prefill and ~35 t/s decode with MTP tool calling.
A detailed report on optimizing a production vLLM serving configuration on NVIDIA's DGX Spark, correcting flags that were costing 34% MTP acceptance after reviewing 90+ official NVIDIA documents and running a 69-scenario tool evaluation.
A real-time demo shows 16 concurrent users chatting with Qwen3.6-35B on a single DGX Spark, achieving peak 440 tok/s total and 105 tok/s per user using NVFP4 + MTP-3 on vLLM.