GPU memory
Holds model data and working buffers. The GPU model, software backend and available memory all matter.
Local AI hardware · Auckland
Choose the model. Size the machine around the work.
Explore local AI with Strata, Qwen and DeepSeek. Start with your documents, users and response-time needs, then plan the GPU, memory and storage together.
Discuss a Local AI WorkstationPlanning guide · reviewed 12 November 2026
The configurations below are Xtreme Systems planning examples informed by upstream documentation. They have not been benchmarked by Xtreme Systems. Exact model files, engine version, operating system and workload must be confirmed before supply.
Start with a use case
A local assistant for the work inside your business.
Possible pilots include drafting from internal templates, searching SOPs, summarising approved documents and assisting developers. Test answer quality and staff review time as well as speed.
System RAM
Provides another memory tier and room for your operating system, documents and business applications.
NVMe storage
Stores model files and supports offloaded data. Capacity and sustained read performance both affect the experience.
Context and users
Longer inputs and concurrent sessions change capacity needs. A single-user demonstration does not establish team capacity.
Reference configurations
Three starting points for a scoped pilot.
Storage figures below are proposed drive capacities, not model download sizes. Leave space for model preparation, updates, documents and backups.
PLANNING EXAMPLE · NOT OUR BENCHMARK
Qwen pilot workstation
12–16 GB GPU · 64 GB RAM · 1 TB NVMePlanning example: a supported NVIDIA RTX GPU with Qwen3.8-Flash-Next Q2_0 or IQ2_XS. Begin with one user and short business tasks; measure your actual document and chat workloads.
Read upstream requirements ↗PLANNING EXAMPLE · NOT OUR BENCHMARK
Qwen with more headroom
24–32 GB GPU · 96–128 GB RAM · 2 TB NVMePlanning example for comparing larger quantisations, document workloads and other apps running alongside the model. More capacity is not a promise of a particular speed or user count.
Read upstream requirements ↗PLANNING EXAMPLE · NOT OUR BENCHMARK
DeepSeek evaluation workstation
12 GB+ supported GPU · 64 GB RAM · 1 TB NVMePlanning example for DeepSeek V4.1-Flash-1017, ISTA-DAS 2.00 bpw. Its approximately 257 GB download includes shared tables and the drafter. SSD offload makes this an evaluation starting point; validate response time before choosing it for daily team use.
Read upstream requirements ↗Model-specific requirements
Keep model size, RAM and VRAM separate.
Qwen3.8-Flash-Next
For the original ISTA-DAS packs, Strata lists roughly 34–50 GB of experts in RAM, depending on quantisation. Its guide allows additional room for the system: 48 GB RAM fits the smaller Q2_0 and IQ2_XS options; 64 GB accommodates the listed original packs, with limited spare capacity at the largest size. Larger Unsloth files need a separate memory calculation.
DeepSeek V4.1-Flash-1017
The standard 1.75, 2.00 and 2.25 bpw sets require approximately 239, 257 and 274 GB of downloads respectively. These totals include the shared 111 GB Engram table and 8 GB draft model. Download size is not the amount of VRAM required: the engine distributes data across the GPU, RAM and SSD.
Before choosing a system
Measure the experience your team will actually have.
Confirm compatibility
Check the exact GPU, OS, driver, engine and model pack together. Qwen and DeepSeek do not share every supported hardware path.
Run a representative pilot
Use approved sample documents. Record time to first answer, output speed, accuracy, context size and peak memory.
Plan the complete machine
Check CPU, memory bandwidth, cooling, power, expansion slots and NVMe performance before buying additional GPUs.
Decide who operates it
Assign model updates, access management, backups, support and review responsibilities before a team rollout.
Local deployment and privacy
Keep control of where information goes.
Local inference can keep model processing on your machine. Web tools, external model services, remote connectors and shared server access introduce separate data flows. Review those settings before using business information.
Strata is a third-party project; Xtreme Systems does not claim a vendor partnership. Model licences are separate from the engine licence. We can help scope hardware procurement and an implementation pilot around your requirements.
Strata privacy and network behaviour ↗ · Our responsible AI approach · Watch: RAG and enterprise knowledge bases
Let’s talk
Tell us what you want to run locally.
Include your current PC, GPU and VRAM, RAM, operating system, target model, typical documents and number of users. We’ll help scope a practical configuration and pilot.