AI & Home Lab

ASUS GX10 & GB10 local models

Two ASUS Ascent GX10 nodes powered by the NVIDIA GB10 Grace Blackwell Superchip — 128 GB unified memory per node, ~1 PFLOP FP4 each, and a private Ollama stack for agent-grade inference on-premises.

NVIDIA GB10 Grace Blackwell 128 GB LPDDR5x NVLink-C2C ConnectX-7 @ 200 Gbps DGX OS Wi-Fi 7

Two ASUS Ascent GX10 nodes at the published 150 by 150 by 51 mm. Stellar grey, with the stacked-wave lid from the iF design note. Drag to turn them. This is a sized study, not a product photo.

DGX Spark beside a DGX Station, same scale. Spark is 150 by 150 by 50.5 mm with the charcoal body and champagne lid from NVIDIA’s product render. The tower is 232 by 569 by 611 mm, the Dell Pro Max chassis. NVIDIA does not publish one Station shell.

NVIDIA GB10 Grace Blackwell Superchip

The GB10 integrates a Grace Arm CPU and Blackwell GPU on one coherent memory domain. CPU and GPU share 128 GB LPDDR5x via NVLink-C2C — about 5× the bandwidth of PCIe 5.0 — so large models stay resident without constant host/device copies.

Compute

Architecture	NVIDIA Grace Blackwell
CPU	20-core Arm v9.2-A (10× Cortex-X925 + 10× Cortex-A725)
GPU	Blackwell generation (integrated)
CUDA cores	Blackwell generation
Tensor cores	5th generation (FP4 / FP8 / FP16)
RT cores	4th generation
AI performance	Up to ~1 PFLOP (FP4 tensor)
SOC TDP	140 W (240 W system peak)

Memory & interconnect

System memory	128 GB LPDDR5x coherent unified
Memory interface	256-bit
Memory bandwidth	Up to 273 GB/s (8533 MT/s)
CPU ↔ GPU link	NVLink-C2C (coherent)
Model capacity	Fine-tune / infer up to ~200B params (single node)
Dual-node scale	2 PFLOP FP4 · 256 GB unified · via ConnectX-7 QSFP
Cluster scale	4+ nodes over network switch (ASUS-supported)

ASUS Ascent GX10 hardware

A 1.48 kg desktop AI supercomputer — 150 × 150 × 51 mm — with enterprise I/O, MIL-STD-810H rugged design, and a full NVIDIA AI software stack preloaded on DGX OS.

Storage

1 TB / 2 TB M.2 2242 NVMe PCIe 4.0 ×4
4 TB M.2 2242 NVMe PCIe 5.0 ×4 option
Single SSD slot (factory sealed chassis)

Networking

1× RJ-45 10 GbE LAN
1× NVIDIA ConnectX-7 SmartNIC (200 Gbps QSFP)
Wi-Fi 7 (2×2) · Bluetooth 5.4 LE
Multi-node over QSFP / switch fabric

I/O & power

3× USB 3.2 Gen 2×2 Type-C (20 Gbps, DP alt)
1× USB-C PD in (EPR PD3.1, 180–240 W)
1× HDMI 2.1a · Kensington lock
240 W external adapter · 1× NVENC / 1× NVDEC

Full platform summary

Spec	ASUS Ascent GX10
Model / SKU family	ASUS Ascent GX10 (GB10 Grace Blackwell)
Dimensions (W × D × H)	150 × 150 × 51 mm (5.91 × 5.91 × 2.01 in)
Weight	1.48 kg (3.26 lb)
Operating system	NVIDIA DGX OS (Linux-based, factory configured)
AI software stack	CUDA · cuDNN · TensorRT · PyTorch · TensorFlow · Jupyter · NVIDIA NIMs
Supported model families	Llama 3.x · DeepSeek R1 · Qwen · Gemma · Nemotron · Meta / Google frameworks
Certifications	BSMI · CB · CE · FCC · UL · CCC · Wi-Fi · MIL-STD-810H
Typical use cases	Local LLM inference · agentic AI · fine-tuning · RAG · ComfyUI image pipelines

Home lab fleet

gx10-eed4 · primary dev · Ollama · Hannah + Nicole agents

gx10-46ab · secondary node · overflow inference

Private mesh · ComfyUI · Cursor agents · no public endpoints

Why GB10 on the desk

→ 128 GB unified memory holds 27B–70B models quantized on one node
→ FP4 tensor path for fast agent loops and tool calling
→ Data stays local — no cloud inference for family or business drafts
→ ConnectX-7 ready when a second GX10 doubles capacity to 256 GB

Local models & capabilities

Pulled from the live Ollama inventory on gx10-eed4. All workloads are business-appropriate and family-safe.

Qwen3.8 27B

27.3B params · Q4_K_M · ~17.7 GB on disk

Primary

General-purpose agent model with tool use, chain-of-thought reasoning, and vision input. Used for drafting, code assist, and multi-step Hermes agent tasks.

Completion Tools Thinking Vision

Context: 262,144 tokens
Embedding dim: 5,120
Best for: Agents · long docs · analysis

Gemma 4 31B

31.3B params · Q4_K_M · ~19.9 GB on disk

Active

Google's Gemma 4 line — strong instruction following and structured output. Good for summaries, email drafts, and lightweight tool-calling workflows.

Completion Tools Thinking

Format: GGUF
Family: gemma4
Best for: Summaries · writing · QA

Nemotron 3 Nano 30B

31.6B params · Q4_K_M · ~24.3 GB on disk · MoE

Active

NVIDIA Nemotron hybrid MoE — optimized for efficient inference on Grace Blackwell. Excels at long-context retrieval, reasoning chains, and tool orchestration with lower active parameter cost.

Completion Tools Thinking

Context: 1,048,576 tokens
Embedding dim: 2,688
Best for: RAG · mega-context · research

Previously hosted & rotation pool

Llama 3.x (8B–70B)

General inference · chat · instruction

DeepSeek R1 (distilled)

Multi-step reasoning · math · planning

Mistral / Mixtral 8×7B

MoE chat · retired from active slot

Phi-3 / Phi-4

Small fast models · edge tests

CodeLlama / StarCoder2

Code completion · refactor assist

nomic-embed-text

Embeddings · RAG indexing

SDXL / Flux (ComfyUI)

Image generation · Krea-2 Turbo pipelines

Llama 3.2 Vision

Image + text · document OCR tests

Qwen 2.5 Coder

Code agents · CI log analysis

Public hardware overview only — no Tailscale IPs, API keys, or internal endpoints. Live model list synced from Ollama on gx10-eed4.