Two NVIDIA GB10 nodes
A connected two-node environment for models requiring more shared memory and compute.
DUAL GB10 · LOCAL AI · LONG CONTEXT
An AI lab
where models run locally.
Two NVIDIA GB10 compute nodes form an experimental environment for substantive analysis, engineering, RAG and testing new AI workflows. Local models are used when they fit the task—not simply for the sake of running locally.
Purpose
More control over compute, data and experiment cost
The lab can process large document corpora and long contexts without sending every request to an external provider. Important outputs still pass a separate quality gate, while a genuinely difficult stage can be deliberately escalated to a stronger cloud route.
Working environment
What the laboratory contains
DeepSeek V4 Flash
The primary local route for substantive analysis, research and complex drafts.
Up to 1,048,576 tokens
The verified configuration can work with large source sets without an artificial few-thousand-token window.
QMD and RAG
Retrieval finds relevant documents and decisions first, then supplies the model with context and sources.
The task chooses the model
Extraction, analysis, coding, critique and visual work are separated and routed to suitable workers.
Tests and quality gates
Results are judged by facts, completeness, reproducibility and practical value—not by confident wording.
Workflow
From task to accepted result
Applied task
The observable outcome comes before the choice of model or technology.
Sources and memory
RAG retrieves the smallest sufficient, verifiable context.
Local route
DeepSeek or a specialist model performs a bounded stage.
Test or expert review
A weak result is repaired or deliberately escalated.
Practical lessons
A local model is part of a system, not the goal
Context must be real — a large advertised window is useless if the wrapper truncates input or output.
Stages should be separated — extracting facts, proposing, implementing and criticising require different assignments.
A routing defect is not model weakness — parameters, limits, finish reason and actual delivered context are checked first.
Accepted work matters — speed and savings count only when the result is good enough.
Related topic
Compute becomes useful through shared memory
The next architectural layer stores decisions, documents and retrieval indexes for primary and backup working environments.
How shared memory for AI agents works →