whollama

Find the best Ollama model for your hardware, ranked by real benchmarks.

Auto-detect your GPU/CPU/RAM and get ranked recommendations from the Ollama library — scored by real benchmark data, not just parameter count.

Quick Start

npx @tsany/whollama

That's it. whollama detects your hardware, fetches the latest model catalog and benchmark scores, and shows you the top recommendations.

Install

npm install -g @tsany/whollama   # global install
# or
npx @tsany/whollama              # zero-install, always latest
# or (macOS)
brew tap tsanys/whollama
brew install whollama

Usage

whollama                  # Show top 10 recommended models
whollama --top 5          # Show top 5 (1-100)
whollama --task coding    # Filter by task (coding, vision, math, tools, embedding, general)
whollama --json           # JSON output for scripting
whollama --offline        # Force offline mode (bundled catalog + scores)
whollama --verbose        # Show scoring breakdown
whollama --no-color       # Disable colored output

whollama --gpu "RTX 4090" --ram 32   # Simulate hardware (planning)
whollama --gpu "M2 Max" --vram 32

whollama pull             # Interactive model pull (fuzzy filter)
whollama pull qwen3:14b   # Pull a specific model

whollama list             # List all fitting models
whollama list --all       # List all models (including non-fitting)

whollama info qwen3:14b   # Detailed model info

whollama bench            # Benchmark smallest pulled model, calibrate speeds
whollama bench qwen3:14b  # Benchmark a specific pulled model

whollama update           # Force refresh catalog and benchmarks

Features

• Hardware auto-detection — Apple Silicon, NVIDIA, AMD, or CPU-only
• Live catalog — scrapes ollama.com/library for the latest models
• Live benchmarks — LiveBench dated tables, merged with a curated snapshot for full coverage
• Smart scoring — composite score factoring benchmark quality, VRAM fit, speed, and recency
• Offline mode — bundled fallback catalog and scores work without internet
• JSON output — pipe-friendly for scripts and automation
• Zero config — no API keys, no setup, no accounts

How It Works

1. Detect — GPU, VRAM, RAM, disk, and OS are auto-detected
2. Fetch — Model catalog from Ollama + benchmark scores from multiple sources
3. Score — Each model is ranked by a weighted composite of benchmark quality, VRAM fit, estimated speed, and recency
4. Display — Top recommendations shown in a formatted table, or as JSON

Output

╭─────────────────────────────── Hardware ───────────────────────────────╮
│ GPU: Apple M1 Pro — 16 GB unified  BW: 200 GB/s                        │
│ RAM: 16 GB  •  Disk: 450 GB free                                        │
╰─────────────────────────────────────────────────────────────────────────╯

  Recommended Models (task: general)

  #   Model               Params   Quant     Score   Speed     Tags
  1   qwen3:14b           14.8B    Q4_K_M    87.4    22 t/s    tools
  2   gemma4:12b          12.0B    Q4_K_M    84.1    28 t/s    vision
  3   llama3.2:3b          3.2B    Q8_0      71.2    65 t/s    tools
  ...

  Top pick: qwen3:14b  •  Run: ollama pull qwen3:14b

Why whollama?

Running local LLMs via Ollama requires answering two questions that are hard to answer together:

1. What can my hardware run? — VRAM, RAM, and disk constrain viable models
2. Which of those is actually the best? — Parameter count is a poor proxy for quality

whollama answers both in a single command.

Offline & caching

No network? No problem — whollama ships a bundled catalog and benchmark scores:

| Cache file | TTL | Contents | |---|---|---| | ~/.whollama/catalog.json | 24 hours | Scraped Ollama library models | | ~/.whollama/benchmarks.json | 7 days | Merged benchmark scores |

whollama --offline skips all network requests and uses the cache (or the bundle on first run). Tiny or corrupt caches are ignored automatically and fall back to the next source.

Speed calibration

Estimates are based on memory bandwidth heuristics. whollama bench measures your real tokens/sec on a pulled model (via the local Ollama daemon) and saves a calibration ratio to ~/.whollama/config.json — future recommendations scale their speed estimates by it. Re-run after hardware or driver changes.

Requirements & troubleshooting

• Node.js ≥ 20 (CI-tested on 20 and 22).
• Ollama must be installed for whollama pull — see ollama.com. Without it, recommendation commands (default, list, info) still work; only pull fails with Failed to pull model.
• No models match? Your hardware filter excluded everything — try whollama list --all or simulate with --gpu "RTX 4090".

License

MIT