Recommended Models

NGIBS is model-agnostic, but performance is hardware-dependent. Choose the right Ollama model for your machine.

Because NGIBS orchestrates complex reasoning loops (especially in Deep Search mode), the underlying LLM needs to be competent at tool-use, context aggregation, and logic. A model that is too small might hallucinate during research; a model that is too large will bottleneck your system.

VRAM is King: For the best generation speeds, your chosen model should fit entirely within your GPU's VRAM. If it spills over into system RAM, token generation will slow down significantly.

Hardware Tiers

Tier 1: 8GB RAM (Laptops & Older PCs)

Stick to lightweight models (under 9 Billion parameters). They are fast, responsive, and take up very little space, but they may occasionally struggle with the heaviest Deep Search recursive reasoning.

Llama 3.1 (8B) Generalist

The gold standard for small models. Unbelievably smart for its size and perfectly capable of handling Live Search scraping.

ollama run llama3.1:8b

Phi-3 Mini Code / Logic

Microsoft's tiny model punches way above its weight class. Very fast, ideal for Quick Search programming questions.

ollama run phi3

Tier 2: 16GB - 32GB RAM (Modern Workstations)

The sweet spot for NGIBS. These models have excellent multi-step context aggregation and handle agentic workflows brilliantly.

Mistral NeMo (12B) Highly Recommended

Built jointly by Mistral and NVIDIA, this model has a massive 128k context window, making it the absolute best choice for reading massive documents and scraping long web pages.

ollama run mistral-nemo

Command-R (35B) RAG Specialist

Cohere designed this model specifically for Retrieval-Augmented Generation (RAG) and tool use. If you have 32GB of RAM, this is the ultimate Deep Search engine.

ollama run command-r

Tier 3: 64GB+ RAM / Multiple GPUs

For running massive models locally to achieve unparalleled reasoning. Expect GPT-4 class logic running entirely offline.

Llama 3.1 (70B) Enterprise Logic

Massive, brilliant, and highly capable. It will rarely make logical errors during Deep Search compilation.

ollama run llama3.1:70b

Qwen 2.5 (72B) Heavy Duty

Incredible coding and multi-lingual capabilities. A behemoth that requires serious hardware.

ollama run qwen2.5:72b

Embedding Models (For Vector Memory)

If you plan to use the Context Aware mode or upload File Attachments, NGIBS uses ChromaDB to store vector embeddings. You need a dedicated, tiny embedding model to convert your text into math.

Nomic Embed Text Default

A highly efficient embedding model that maps text locally. Fast, accurate, and requires almost zero resources.

ollama pull nomic-embed-text

Changing Models in NGIBS

You do not need to use the terminal to switch models once Ollama is installed.

  1. Open NGIBS.
  2. Navigate to Settings > LLM Manager.
  3. To download a new model, type its registry name (e.g., mistral-nemo) in the text box and click Pull Model. NGIBS will display a progress bar as it downloads.
  4. To switch the active model, use the dropdown menu at the top of your chat window. You can swap models mid-conversation seamlessly.