Ollama Model Manager Skill

Overview

Manage local Ollama models: installation, removal, benchmarking, comparison, and recommendations for different tasks.

Prerequisites

• Ollama running: ollama serve (or service running)
• Default host: http://localhost:11434
• Alternative: Set OLLAMA_HOST env var

Phase 1: STATUS

Check installed models and system state:

List Models

ollama list
# Output:
# NAME                       ID              SIZE      MODIFIED
# qwen3-coder:latest         abc123...       4.7 GB    3 hours ago
# llama3:latest              def456...       4.7 GB    1 day ago
# nomic-embed-text:latest    ghi789...       274 MB    5 days ago

Detailed Model Info

ollama show qwen3-coder
# Output:
# Model
# architecture       qwen2
# parameters         27b
# quantization       Q4_K_M
# context length     32768
# Tags
# modelfile          /Users/.../Modelfile
# parameters         /Users/.../parameters

Check Ollama Status

curl -s http://localhost:11434/api/tags
# Returns JSON: {models: [{name, size, modified_at}]}

Running Models (Active)

ollama ps
# Shows currently loaded models in memory

Phase 2: SEARCH

Find and evaluate available models:

Browse Available Models

https://ollama.com/library

Model Categories

Code-Focused

• qwen3-coder (27B) — best overall code assistant
• deepseek-coder (33B, 6.7B) — specialized coding
• codellama (34B, 13B, 7B) — Meta's code LLM
• phi3-code — lightweight code model

General Purpose

• llama3 (70B, 8B) — high quality, fast
• mistral (7B) — efficient, good quality
• mixtral (46.7B, 8x7B) — mixture-of-experts
• neural-chat (7B) — conversational

Embeddings/Search

• nomic-embed-text (274M) — fast, good quality
• mxbai-embed-large (334M) — better quality

Specialized

• phi (2.7B) — very small, surprisingly capable
• orca-mini (3B) — small but smart
• mistral-nemo (12B) — good balance
• openchat (7B) — chat optimized

Pull Model

ollama pull qwen3-coder
# Downloads to ~/.ollama/models/ (or OLLAMA_MODELS path)
# Size: ~4.7GB for qwen3-coder Q4 variant

Pull Specific Version/Quantization

ollama pull llama3:70b         # 70B version
ollama pull llama3:7b          # 7B version (faster, smaller)
ollama pull phi3:medium        # Medium variant
ollama pull nomic-embed-text   # Embedding model

Phase 3: BENCHMARK

Compare model performance on same task:

Single Model Benchmark

time ollama run qwen3-coder "write a python quicksort function"
# Measures: real time (wall clock), user time, system time

Detailed Benchmark with Timing

ollama run qwen3-coder --verbose "prompt" 2>&1 | tee bench.log
# Verbose shows token timing and generation speed

Benchmark Script (multiple models)

#!/bin/bash

PROMPT="Write a Python function that reverses a list without using slicing."
MODELS=("qwen3-coder" "llama3:8b" "codellama")

echo "Model,Time(s),Tokens,Tokens/Sec,Quality(1-10)"

for model in "${MODELS[@]}"; do
  start=$(date +%s%N)
  output=$(ollama run "$model" "$PROMPT" 2>&1)
  end=$(date +%s%N)
  
  elapsed=$(( (end - start) / 1000000 ))
  elapsed_sec=$(echo "scale=2; $elapsed / 1000" | bc)
  
  # Rough token count (1 token ~= 4 chars on average)
  tokens=$(( ${#output} / 4 ))
  tokens_per_sec=$(echo "scale=2; $tokens / $elapsed_sec" | bc)
  
  # Manual quality rating (1-10)
  echo "$model,$elapsed_sec,$tokens,$tokens_per_sec,?"
done

Comparison Table

Create markdown table comparing:

• Model name
• Size (GB)
• Speed (tokens/sec)
• Quality (subjective 1-10)
• Best for (coding, general, chat, embeddings)
• Memory requirement
• Quantization (Q4, Q5, FP16)

Example:

| Model | Size | Speed | Quality | Best For | Memory |
|-------|------|-------|---------|----------|--------|
| qwen3-coder | 4.7GB | 25 tok/s | 9/10 | Code | 8GB |
| llama3:70b | 39GB | 8 tok/s | 8/10 | General | 48GB |
| llama3:8b | 4.7GB | 40 tok/s | 7/10 | Chat | 8GB |
| codellama | 34GB | 12 tok/s | 8/10 | Code | 40GB |
| phi3 | 2.3GB | 80 tok/s | 6/10 | Fast | 4GB |

Phase 4: MANAGE

Install, remove, and configure models:

Pull Model (Download)

ollama pull qwen3-coder
# Downloads latest version with default quantization
# Typically Q4_K_M (good balance of speed/quality)

Remove Model

ollama rm qwen3-coder
# Permanently deletes local model (no recovery)
# Always confirm before running

Create Custom Model (Modelfile)

FROM llama3:70b
PARAMETER temperature 0.7
PARAMETER top_k 40
PARAMETER top_p 0.9
SYSTEM """You are an expert code reviewer..."""

Then:

ollama create custom-reviewer -f Modelfile

Export Model

# Models stored in ~/.ollama/models/blobs/
# Copy manually to backup location

Check Disk Space Used

du -sh ~/.ollama/models/
# Example output: 45G

# Per model:
ls -lh ~/.ollama/models/blobs/ | head -10

Phase 5: RECOMMEND

Suggest best model for task type:

Code Development & Review

• Primary: qwen3-coder (best instruction following, code quality)
• Alternative: deepseek-coder (strong coding, fast)
• Fast: codellama:7b (quick, decent quality)
• Size constraint (<2GB): phi3-code or orca-mini

General Chat & Writing

• High quality: llama3:70b (best but slow)
• Balanced: llama3:8b (good speed/quality)
• Fast: mistral:7b
• Very fast: phi3

Data Analysis & Reasoning

• Best: llama3:70b (larger models reason better)
• Medium: mixtral:46.7b
• Fast: llama3:8b

Embeddings & Search

• Standard: nomic-embed-text (274M, fast, good)
• Quality: mxbai-embed-large (334M, better quality)

Production Retrieval-Augmented Generation (RAG)

• Embedding: nomic-embed-text
• LLM: qwen3-coder or llama3:8b
• Keep in memory: yes (fast inference)

When Disk/Memory Limited

Available RAM       Recommended Model
8GB                 llama3:8b, phi3, mistral:7b
12GB                qwen3-coder, codellama:7b
16GB+               llama3:70b, mixtral

Decision Tree

1. Is it coding? → qwen3-coder
2. Need reasoning? → llama3:70b or mixtral
3. Speed critical? → phi3 or mistral:7b
4. Embeddings? → nomic-embed-text
5. Default → llama3:8b

Common Workflows

Setup for Development

# Core models for coding projects
ollama pull qwen3-coder
ollama pull nomic-embed-text

# Optional: general chat
ollama pull llama3:8b

# Verify
ollama list

Benchmark Before Deciding

# Download 2-3 candidate models
ollama pull qwen3-coder
ollama pull codellama:7b
ollama pull phi3

# Run benchmark script
./benchmark.sh

# Choose based on results
ollama rm codellama  # if slower

Switch Active Model for Task

# In your application, specify model:
ollama run qwen3-coder "code task"
ollama run llama3:8b "general task"

Monitor Memory Usage

ollama ps    # shows loaded models
top          # watch memory growth

# If memory exhausted, unload models:
ollama ps    # see what's loaded
# Models unload automatically after 5 min idle
# Force unload by restarting ollama service

Cleanup Old Models

ollama list
ollama rm old-model-name
# Frees disk space for new models

Troubleshooting

"Connection refused"

• Check Ollama is running: ollama serve
• Check host: OLLAMA_HOST=0.0.0.0:11434 ollama serve (if remote)

Model download stuck

• Network issue: restart and retry
• Disk full: check df -h, clean up

Very slow inference

• Model larger than RAM
• GPU not available
• System under heavy load

Memory leak / excessive growth

• Restart ollama service
• Smaller model may fit better