tinillm 1.5.0

Hardware LLM capability scanner — know what runs on your machine

tinillm

Know what LLMs your hardware can run — locally, instantly.

pip install tinillm
tinillm scan

---

What it does

tinillm scan inspects your CPU, RAM, and GPU then tells you exactly which LLM model sizes can run on your machine, at what quality level, and how fast.

╭──────────────────────────────────────────────────────╮
│   tinillm scan — Hardware Report                     │
├──────────────────────────────────────────────────────┤
│  CPU    Intel Core i9-13900K   24c / 32t   5.8 GHz   │
│  RAM    32.0 GB total  ·  24.2 GB free               │
│  GPU    NVIDIA GeForce RTX 4090   24.0 GB   CUDA     │
│  OS     Linux                                        │
╰──────────────────────────────────────────────────────╯

  LLM Capability Matrix

  Model    Fit        Best Quant   Mem Needed   Tokens/sec
  ~1B      Perfect    Q8_0          1.9 GB        580 t/s
  ~3B      Perfect    Q8_0          3.8 GB        195 t/s
  ~7B      Perfect    Q6_K          6.2 GB         88 t/s
  ~13B     Perfect    Q5_K_M       10.1 GB         47 t/s
  ~34B     Good       Q4_K_M       21.8 GB         18 t/s

  1 model size(s) too large — use --verbose to show
  Perfect  ·  Good  ·  Marginal  ·  TooTight

Works on Linux, macOS, and Windows. No GPU required — CPU-only machines are supported too.

---

Install

pip install tinillm

Requires Python 3.11+. No other tools needed.

---

Usage

tinillm scan                # hardware + LLM capability report (default)
tinillm scan --verbose      # show all model sizes including ones that don't fit
tinillm scan --json         # machine-readable JSON (for scripts / CI)
tinillm scan --no-color     # plain text (safe to pipe to grep / awk / log files)

Scripting example

# Find all models that run perfectly on this machine
tinillm scan --json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for fit in data['fits']:
    if fit['fit_level'] == 'Perfect':
        print(fit['model'], fit['best_quant'], fit['tokens_per_sec'], 't/s')
"

---

GPU support

Vendor	Detection method
NVIDIA	nvidia-smi → sysfs fallback
AMD	rocm-smi → sysfs fallback
Apple Silicon	system_profiler (unified memory)
Intel Arc	sysfs + lspci
Windows (all)	PowerShell WMI
Any	vulkaninfo last-resort fallback

---

Fit levels explained

Level	Meaning
Perfect	Fits comfortably at Q4_K_M or better with ≥20% VRAM headroom
Good	Fits but tightly
Marginal	Fits only at heavy compression / reduced context, or CPU-only
TooTight	Won't fit under any quantisation

---

Versioning

tinillm grows one feature at a time:

Version	Feature
1.1	scan — hardware LLM capability scanner ← current
1.2	(next feature)

---

Part of the tini* family

Tool	What it does
tiniRAG	Privacy-first RAG CLI
tinillm	Hardware LLM capability scanner