NestAI Docs

AI MODELS

NestAI runs open-source AI models directly on your server using Ollama. You control which models are installed — and your data never leaves your infrastructure. All speeds below are for a single user on CPU. Concurrent users share CPU and will see lower speeds.

ℹ

Models are grouped by tier. Your standard server (8 vCPU, 16GB RAM, 160GB SSD) runs most models well. Add dedicated resources from the billing page for 32B+ models.

Latest & Recommended Models (2026)

These models run well on standard and dedicated servers. Best balance of speed, quality, and reliability.

Model	Size	Best for	Speed (single user)
Qwen 3.5 4B ⭐	2.8 GB	Chat, vision, multilingual	~15-20 tok/s
Qwen 3.5 9B	5.5 GB	All-round, vision, reasoning	~8-12 tok/s
Llama 3.2 3B	2.0 GB	Fast chat, Q&A, drafting	~15-25 tok/s
Phi-4 Mini	2.5 GB	Reasoning, beats most 7B	~15-20 tok/s
Gemma 3 4B	3.3 GB	Multilingual, instruction	~12-18 tok/s
Llama 3.1 8B	4.7 GB	Balanced general use	~8-12 tok/s
DeepSeek R1 7B	4.7 GB	Reasoning, math, logic	~10-14 tok/s
Mistral 7B	4.1 GB	Business, code, structured	~10-15 tok/s
Qwen 2.5 Coder 7B	4.5 GB	Coding	~10-14 tok/s
Gemma 3 12B	7.5 GB	High-quality all-round	~6-8 tok/s
Mistral Nemo 12B	7.1 GB	Long documents (128K ctx)	~6-8 tok/s

ℹ

Qwen 3.5 4B is our top recommendation for most users — it outperforms many 7B models while being faster and smaller.

Advanced Models (14B-20B)

Higher quality outputs but slower on standard servers. Best with dedicated resources.

Model	Size	Best for	Speed
GPT-OSS 20B 🆕	12 GB	OpenAI open-weight, adjustable reasoning	~3-5 tok/s
Phi-4 14B	8.9 GB	STEM, reasoning (GPT-4o-mini level)	~4-6 tok/s
Phi-4 Reasoning 14B 🆕	9.0 GB	Math olympiad, complex logic	~4-6 tok/s
DeepSeek R1 14B	9.0 GB	Advanced math & logic	~4-6 tok/s
Qwen 2.5 14B	8.5 GB	Multilingual, analysis	~3-5 tok/s
Qwen 2.5 Coder 14B	9.0 GB	Complex coding, multi-file	~4-6 tok/s

⚠

These models fit in 16GB RAM but are significantly slower on shared servers (~3-6 tok/s). Add dedicated resources for a better experience.

Flagship Models (Dedicated Resources Required)

These models deliver the highest quality but require dedicated server resources (32GB+ RAM). Add dedicated resources from Dashboard → Billing.

Model	Size	Best for	Speed on CPU
DeepSeek R1 32B	19 GB	Near GPT-4 reasoning	~4-6 tok/s
Qwen 2.5 32B	19 GB	Multilingual + general	~4-6 tok/s
Qwen 2.5 Coder 32B	19 GB	Best coding model	~4-6 tok/s
Llama 3.3 70B	43 GB	Highest quality overall	~2-3 tok/s

ℹ

70B models run at ~2-3 tokens/sec on CPU — usable for document analysis and batch tasks, but not ideal for real-time chat. 32B models are the sweet spot for quality vs speed.

Embeddings & RAG

Model	Size	Purpose
Nomic Embed Text	274 MB	Required for Knowledge Base (RAG) and document search

✓

Install Nomic Embed to enable document search, semantic retrieval, and knowledge base features.

Installing Models

Go to Dashboard → Models and click Pull →. Downloads run in the background and typically take 2–15 minutes depending on model size. You can install multiple models and switch between them anytime.

⚠

Do not restart your server while a model is downloading.

Removing Models

Click Remove next to any installed model to free disk space. You can reinstall models anytime.

Disk Usage

Setup	Disk Usage
1-2 small models (3-4B)	~5-8 GB
2-3 models (7B)	~10-15 GB
5 models mixed	~25 GB
System + base setup	~8 GB

ℹ

Standard servers include 160GB SSD. Dedicated resource tiers include 240-960GB depending on tier.