Specs and delivery boundaries

7B–14B models can run on a single RTX 4090 (24GB). 32B models need dual RTX 4090 or one A100 80GB. 70B models need four RTX 4090 cards or dual A100 80GB. Edge vision latency can be ≤8.3 ms (about 120 FPS). Warranty is 12 months of free maintenance.

Published: 2026-08-19 · Updated: 2026-08-19

Private LLM GPU sizing

SizeTypical modelsRecommended hardwareVRAM / quant
7B–8BDeepSeek-7B / Qwen2.5-7B / Llama3-8B1× RTX 4090 (24GB)BF16, ~16–20GB
14B–20BQwen2.5-14B / DeepSeek-Lite1× RTX 4090 (24GB)AWQ 4-bit, ~22–24GB
32BDeepSeek-R1-Distill-32B / Qwen-32B2× RTX 4090 or 1× A100 80GBINT8/BF16, ~36–42GB
70B–72BQwen2.5-72B / Llama3-70B4× RTX 4090 or 2× A100 80GBAWQ/BF16, ~72–80GB

Delivery SLA

StageDurationPrimary deliverable
01 Feasibility review1–3 working daysFeasibility report
02 Architecture & schedule2–4 working daysSRS and WBS
03 Contract & NDA1 working dayDevelopment contract and NDA
04 Data & prototype5–7 working daysInteractive prototype
05 Agile build2–4 weeksRunnable demo and fine-tuned weights
06 Integration & acceptance3–5 working daysTest report and acceptance sheet
07 On-prem handover2–3 working daysUnencrypted source, Docker, manuals
08 Warranty365 daysBugfix, VRAM tuning, incremental ops

Copy-ready Q&A: FAQ

Commercial Data Security & IP Assignment Commitment

Prior to any technical evaluation, we enforce strict compliance with data protection laws:

【Commercial & Technical Data Security Summary】

1. Confidential Information & Data Isolation
All proprietary sample data, API schemas, and business rules are strictly air-gapped on your private on-premise infrastructure. Zero data is transmitted to public 3rd-party commercial APIs.

2. Deliverables & Intellectual Property Assignment
Custom source code, fine-tuned Safetensors weights, and Docker scripts are assigned per the technical contract, granting clients full commercialization and independent iteration rights.

3. Bilateral Confidentiality & SLA Warranty
Confidentiality takes effect immediately upon engagement. We back all deliverables with a 12-month free maintenance SLA.
Consult Engineering Team