Skip to content

Software stack / reviewed Oct 23, 2026

From the hardware to the request.

DeepSeek serves across both GB10 hosts as of October 23. The same-day review covers Studio launchers and the workstation model inventory. Installed software, retained historical runtimes and pending updates keep their own dates.

Layer 01

Hardware

DGX Spark + ASUS GX10
Two 128 GB unified-memory GB10 hosts. The persistent DeepSeek coder uses TensorFold TP2, one rank on each host.Observation / Oct 23, 2026
Mac Studio / M5 Max
48 GB unified memory; 18 CPU / 40 GPU cores. Agent, writer, embeddings and local inference.Observation / Oct 23, 2026
RTX 5090 / Ryzen 9 9950X3D
32 GB GDDR7. Windows host, WSL2 execution, media and utility specialists.Observation / Oct 23, 2026

Layer 02

Drivers & platform

NVIDIA 580.173.02
Still installed on both GB10 hosts. Driver 580.178.04 was available in the October 22 package inventory; OS/kernel/driver maintenance remains pending.Observation / Oct 23, 2026
NVIDIA 617.42 / WSL2
Directly reported by the workstation WSL driver on October 23; kernel 6.18.33.2. The earlier October 2 observation was driver 617.14.Observation / Oct 23, 2026
macOS 27.0.1 / Metal
Verified local OS and M5 Max GPU. Unified-memory limits remain workload dependent.Observation / Oct 23, 2026

Layer 03

Inference runtimes

vLLM 0.30.0 / historical Qwen deployment
October 10 Flash-Next NVFP4 deployment and its eight-slot, 3.87× startup-cache calculation remain historical. These settings and measurements do not describe the active DeepSeek runtime.Observation / Oct 10, 2026
TensorFold native / DeepSeek V4.1 Flash 1017
Active EXL3 EAQ 2.9 bpw checkpoint, normal mode, TP2, four shared slots and 524,288 configured context. Recipe 370daf0, engine 629c524, patch set cc4781a03264. High reasoning default; 32,768 output allowance. GLM’s exact rollback resources are preserved.Observation / Oct 23, 2026
MLX / mlx-lm / mlx-vlm
Studio agent alias now serves SuperQwen MLX 4-bit with thinking off and one active request; Qwen is a separate fallback. September evaluated versions: MLX 0.32.2, mlx-lm 0.31.3, mlx-vlm 0.7.2. Nemotron Lightning measurements do not transfer to SuperQwen.Observation / Oct 23, 2026
llama.cpp
Studio Gemma and persistent Qwen embedder; workstation GGUF specialist routes. Evaluated Studio build 10964 is a dated result.Observation / Oct 02, 2026

Layer 04

Routing & lifecycle

llama-swap
Customized v257 gateway on both GB10 hosts. Coder, scout and vision share DeepSeek and four inference slots; Studio agent and embeddings stay independent. Upstream v266 is available; local capacity, affinity, peer-readiness and UI patches must be carried forward before replacement.Observation / Oct 23, 2026

Layer 05

Supporting systems

Media pipelines
Workstation Z-Image variants, Krea 2, Qwen Image 2.1 / Turbo / edit, FLUX.2 klein, Wan, LTX, MiniMax H3, Hunyuan, Lingbot, ACE-Step and MiniMax Music 3. On-demand inventory with legacy matrix sets and per-model TTLs; concurrent heavy pipelines remain unqualified. Studio MFLUX alternatives remain in its exclusive primary group.Observation / Oct 23, 2026
sparkDash 2.70.0
Updated from 2.0.0 on October 22 with local fleet, routing and proxy-mount customizations preserved. Distinct from NVIDIA DGX Dashboard. Local build, type checking and frontend/backend checks passed; only curated results are published here.Observation / Oct 22, 2026
Repo prompt loop
Primary Hermes harness now uses DeepSeek. Four-worker target, four shared inference slots, high-reasoning coding and separate short non-thinking inspection/planning budgets. Restarted and resumed October 23 after regressions; held sessions retained. Draft delivery, independent review and merges remain separate outcomes.Observation / Oct 23, 2026
Pending platform maintenance
Inventory only: DGX Dashboard 0.29.2-2 → 0.39.3-4; Spark OTA 26.04.1 → 26.09.3; Caddy 2.11.4 → 2.11.7; patched Grafana 13.1.2 → 13.2.4; Prometheus 3.13.2 → 3.15.1. Review custom patches and vendor-specific firmware timing before a separate maintenance window.Observation / Oct 22, 2026
How routing and lifecycle work ↗