Store

DeepSeek-V4-Flash on 2× DGX Spark: A Production-TP2 Case Study (and How to Replicate It)

How we serve DeepSeek-V4-Flash — a 671B-parameter MLA-MoE model — across two NVIDIA DGX Spark (GB10) desksides at production quality, with every tuning decision, patch, and failure-mode documented. Powered by the Dual DGX Spark Cluster Blueprint.

Single DGX Spark Recipe: Qwen3.8-27B — Frontier Quality on One Desktop

Run Qwen3.8-27B — a 27B-parameter reasoning model — on a single DGX Spark (GB10) at high throughput with 128K+ context, tool calling, and reasoning. No second node, no RoCE fabric, no speculative-decoding patches. The accessible on-prem frontier.

Knowledge Graph Toolkit

The complete knowledge graph arsenal — GraphRAG pipelines, Neo4j deployment, and monitoring. 330 pages of production-tested guides plus infrastructure.

AI Infrastructure Mastery

The complete blueprint for building a production-grade AI homelab. Connect Dify, n8n, Ollama, Qdrant, monitoring, and backup into one cohesive system.

Dual DGX Spark Cluster Blueprint

Production Ansible deployment for two DGX Spark (GB10) nodes with vLLM TP2 serving DeepSeek-V4-Flash (671B), fronted by LiteLLM, monitored by Prometheus/Grafana, and wired over 200Gbps RoCE. The reference deployment for our Dual DGX Spark case study.

Dify Production Stack

Production-ready Dify with PostgreSQL, Redis, Weaviate, Nginx, and SSRF protection — not a toy compose file.