Skip to content
← All setup guides

Guide 02 / GB10

DGX Spark & GX10: capacity with discipline

Our DGX Spark and ASUS GX10 now run DeepSeek through TensorFold TP2. Understand unified memory, four shared slots, model staging and measured limits.

3 MIN READ / REVIEWED Oct 23, 2026

01Establish the platform

Our DGX Spark and ASUS GX10 each have 128 GB unified memory. Verify your exact SKU before budgeting: NVIDIA now documents 64 GB variants too. Follow vendor setup and update instructions, record the installed driver, and use ARM64 builds with GB10-compatible kernels. Partner firmware follows its own release schedule.

Read-only platform checks

uname -m
nvidia-smi --query-gpu=name,driver_version --format=csv
free -h

02Qualify a single-node baseline

Verify the model, chat template, tool parser, thinking policy and retrieval before adding another node. On our GB10 machines, the GPU total-memory query reports N/A; use system memory and actual process/runtime allocations rather than substituting a fictional VRAM total.

03Separate routing from distributed inference

A gateway selects a backend; tensor parallelism distributes computation. Our current two-node coder is DeepSeek V4.1 Flash 1017 EXL3 on TensorFold, replacing the earlier Qwen/vLLM configuration. Both ranks need matching pinned runtime and model artifacts. The native identity must survive every gateway hop.

Our active configuration

# Inspected deployment policy · 23 Oct 2026
# Inventory excerpt; not a standalone launch command
model: DeepSeek-V4.1-Flash-1017
quantization: EXL3 EAQ 2.9 bpw
runtime: TensorFold native Zig
nodes: 2
tensor_parallel_size: 2
mode: normal
ABLIT: 0
CONTEXT: 524288
PARALLEL: 4
MAX_TOKENS: 32768
THINKING: 1
REASONING_EFFORT: high
MIN_AVAIL_GIB: 112
MEMGUARD_GIB: 2.5
MEMGUARD_HARD_GIB: 1
MEMGUARD_HOLD_S: 4
shared_roles: [coder, scout, vision]

04Treat configured context as a claim to test

DeepSeek has a 524,288-position configured limit and four slots shared by coder, scout and vision. Local validation included a 21,875-position prompt and a separate 63,164-position history replay. Neither is a full-window retrieval or concurrent long-context test. The earlier Qwen 3.87× startup-cache calculation remains attached to the October 10 runtime.

Primary references

Upstream docs describe supported behavior. Lab-specific configurations and numeric results retain their inspection or experiment dates.

Keep going

Related guides