:: SAAT Knowledgebase - Connector Categories ::

Why Private AI?

The Problem	Our Solution
Sensitive data sent to third-party APIs	100% on-premise — your data never leaves your network
Recurring subscription costs	One-time purchase with optional support plans
Compliance headaches (GDPR, HIPAA, etc.)	Air-gapped capable — works without internet
Vendor lock-in	Open-source models — no licensing traps

Package	Best For	Models Included	Hardware Need*	License / Cost
Llama 3.2	General productivity, coding, on-prem privacy	3B – 90B variants	8 GB+ VRAM for 8B; 24 GB+ for 70B	Free (weights); commercial use OK
Qwen 2.5 / Qwen3	Multilingual enterprise apps, long-document analysis	0.5B – 72B variants	6 GB+ VRAM for 7B; 48 GB+ for 72B	Free (Apache 2.0)
Mistral / Mixtral 8x7B	High-throughput APIs, "GPT-4 class" on smaller infra	7B – 8x22B MoE	24 GB+ VRAM for 8x7B; 80 GB+ for 8x22B	Free (Apache 2.0)
Gemma 2	Edge devices, mobile apps, lightweight assistants	2B – 27B variants	4 GB+ VRAM for 2B; 20 GB+ for 27B	Free (Gemma license; some restrictions)
Phi-4 / Phi-3	Cost-sensitive SaaS, student/teacher tools, coding tutors	3.8B – 14B variants	4 GB+ VRAM for 3.8B; 12 GB+ for 14B	Free (MIT)

*Hardware = VRAM needed to run the largest variant comfortably with Ollama; smaller quants need less.

Must keep data on-prem / air-gapped?

Any of the five above work; Llama 3.2 or Qwen for easiest compliance docs.

Primary language is Chinese or need 10+ languages?

Qwen 2.5/3 (best multilingual).

Running on laptops or edge IoT (no GPU)?

Gemma 2 (2B) or Phi-4 (3.8B) quantized.

Need "GPT-4 level" reasoning but only 1–2 A100s?

Mixtral 8x22B or Llama 3.2 70B.

Planning to fine-tune on proprietary small data (<10k samples)?

Phi-4 or Llama 3.2 8B (cheapest training cost).

Redistributing the model inside my own product?

Check license: Phi-4 (MIT) and Qwen (Apache 2.0) are safest; Gemma has extra Google terms.

If you want...	Pick...
Maximum capability, no cloud	Llama 3.2 70B
Global enterprise, long docs	Qwen 2.5 72B
Speed & throughput on mid-range GPUs	Mixtral 8x7B
Cheapest per-token hosting	Phi-4 14B
Mobile / browser / Raspberry Pi	Gemma 2 2B

Analyze confidential documents, contracts, and reports without cloud exposure.

Process patient records and medical literature HIPAA-compliantly.

Air-gapped systems with zero external connectivity for maximum security.

Prototype AI features without API rate limits, costs, or data restrictions.

Unlimited student and faculty access with no per-seat licensing fees.

Internal knowledge bases, SOP generation, and technical documentation.

Contact us for a personalized recommendation based on your infrastructure and use case.