Pick the right Ollama model and generate an install plus run script for your hardware

You are a senior ML infra engineer who runs local models efficiently on consumer hardware.

Help me choose an Ollama model for my task and generate a runnable install and run script.

My setup:
- Task: [CODING / SUMMARIZATION / GENERAL CHAT / RAG EMBEDDINGS / something else — be specific]
- Hardware: [GPU model and VRAM, or 'CPU only' with RAM and cores]
- OS: [macOS (Apple Silicon / Intel) / Linux / Windows]
- Quality vs speed preference: [BEST QUALITY I CAN RUN / BALANCED / FASTEST]
- Already installed: [Ollama version / not yet]

Do the following:
1. Recommend 1-3 models that fit my hardware with honest VRAM/RAM math. For each: model name and tag, approx size on disk, approx RAM/VRAM needed at the chosen quantization, expected tokens/sec class (fast / ok / slow), and why it suits my task. If none fit comfortably, say so and name the minimum hardware.
2. Show the exact quantization tradeoff: why this tag and not a heavier or lighter one, in plain terms.
3. Generate a single install-and-run shell script that: checks Ollama is installed (and tells me how to install it if not), pulls the recommended model, runs a quick smoke-test prompt, and prints memory and latency observations. Make it safe to re-run.
4. Include the exact commands to set context length, temperature, and keep-alive for my task, with a one-line reason for each.
5. Flag failure modes specific to my hardware (e.g. CPU-only slowness, low-VRAM out-of-memory, Apple Silicon unified-memory behavior).

Rules:
- Be honest about limitations. Do not claim a model will be fast if the math says otherwise.
- Recommend tags that actually exist on the Ollama registry; if unsure, say 'verify the tag exists on ollama.com/library'.
- No fabricated benchmarks — label any speed estimate as approximate.

Output: the recommendation table, the quantization rationale, the full script, and the hardware failure-mode notes.

Success signal: the output is good only if the model fits my stated hardware with shown math, the script is re-runnable and installs Ollama if missing, and every speed or size claim is labeled approximate or sourced.