LLM Checker maintains two catalog layers that are merged at runtime: a dynamic catalog scraped from the Ollama registry, and a curated fallback catalog used when the dynamic pool is unavailable.
Dynamic Catalog
When the sync command has been run (requires sql.js), LLM Checker operates against the full scraped Ollama catalog — all families, sizes, and quantization variants. This pool typically covers 200+ models.
# Download the latest catalog
npm install sql.js
llm-checker sync
search and smart-recommend query this database directly with full scoring.
Curated Fallback Catalog
When the dynamic scraped pool is unavailable, LLM Checker falls back to a built-in curated catalog of 35+ models from the most popular Ollama families. This catalog is stored at src/models/catalog.json.
The curated fallback is used only when the dynamic scraped pool is unavailable. If you have run llm-checker sync, the full dynamic catalog is used instead.
| Family | Models | Best For |
|---|
| Qwen 2.5/3 | 7B, 14B, Coder 7B/14B/32B, VL 3B/7B | Coding, general, vision |
| Llama 3.x | 1B, 3B, 8B, Vision 11B | General, chat, multimodal |
| DeepSeek | R1 8B/14B/32B, Coder V2 16B | Reasoning, coding |
| Phi-4 | 14B | Reasoning, math |
| Gemma 2 | 2B, 9B | General, efficient |
| Mistral | 7B, Nemo 12B | Creative, chat |
| CodeLlama | 7B, 13B | Coding |
| LLaVA | 7B, 13B | Vision |
| Embeddings | nomic-embed-text, mxbai-embed-large, bge-m3, all-minilm | RAG, search |
Locally Installed Models
All available catalog models are automatically combined with locally installed Ollama models before scoring. Installed models receive priority consideration in recommendations.
# Rank only your installed models
llm-checker installed
Supported Quantization Types
The following quantization formats are recognized and used for memory estimation and candidate filtering:
| Format | Description |
|---|
Q8_0 | 8-bit quantization — highest quality, ~1.05 bytes/param |
Q4_K_M | 4-bit K-quant medium — best balance, ~0.58 bytes/param |
Q3_K | 3-bit K-quant — smallest footprint, ~0.48 bytes/param |
FP16 | 16-bit float — full precision, largest size |
Q4_0, Q5_0, Q5_K_M | Additional common quantization variants |
llm-checker search qwen --quant Q4_K_M --max-size 8
Fine-Tuning Suitability Labels
check, recommend, and ai-check output include a fine-tuning suitability label for each recommended model:
| Label | Meaning |
|---|
| Full FT | Supports full fine-tuning (requires significant GPU memory) |
| LoRA | Supports LoRA adapter training |
| QLoRA | Supports quantized LoRA (most memory-efficient) |
| LoRA+QLoRA | Supports both LoRA and QLoRA paths |
| Full+LoRA+QLoRA | Supports all fine-tuning modes |
Coding:
qwen2.5-coder:14b (14B)
Score: 78/100
Fine-tuning: LoRA+QLoRA
Command: ollama pull qwen2.5-coder:14b
