Lesson 4.1 — Run Two Models & Pick a Winner

Introduction to AI · Module 4: Local LLMs · About 75 min · Windows PC, RTX 4090, PowerShell 5

Tonight's goal: run qwen3.8:27b and deepseek-r1:32b on 4 of your own real tasks, score them, and choose your everyday local model.

Before you start

• Close Chrome and games (they use GPU memory)
• Water or coffee nearby
• Open Windows PowerShell (Start → type powershell → Enter)

Step 1 · Check your models (⏱ 0:00–0:05)

Type this and press Enter:

ollama list

You should see both models (sizes approximate):

NAME               ID            SIZE     MODIFIED
qwen3.8:27b        ...           18 GB    ...
deepseek-r1:32b    ...           20 GB    ...

• Both models are listed

Then check the GPU:

nvidia-smi

You should see a table with NVIDIA GeForce RTX 4090 and 24564MiB total memory.

• 4090 shows up

Step 2 · First chat and speed check (⏱ 0:05–0:15)

ollama run qwen3.8:27b

You should see a >>> prompt. You're now talking to the model, not PowerShell. Type hi and press Enter. You'll get a reply (it may "think" first).

• Got a reply

Now leave the chat: **type /bye and press Enter. Wait until you see PS C:\...> before the next command.**

• Typed /bye, pressed Enter, and waited until I see PS C:\...> before the next command

Now measure speed (type this at the PS C:\...> prompt, not at >>>). Because the question is included in the command, it answers and returns to PS C:\...> by itself, with no /bye needed:

ollama run qwen3.8:27b --verbose "Explain VRAM in 2 sentences"

You should see the answer, then stats ending with a line like:

eval rate:            40.00 tokens/s

Check that it's on the GPU (run this within 5 minutes, while the model is still loaded):

ollama ps

You should see 100% GPU in the PROCESSOR column.

• 100% GPU shown
• Qwen speed written here: tokens/s

Repeat the speed check for DeepSeek. Switching models happens in the same PowerShell window: no need to open a new one or close anything. Just check you're at PS C:\...>, then run the command with the other model's name:

ollama run deepseek-r1:32b --verbose "Explain VRAM in 2 sentences"
ollama ps

• DeepSeek speed: tokens/s · GPU: %

Step 3 · Set up your 4 tasks (⏱ 0:15–0:25)

Make a test folder:

mkdir C:\llm-test -Force
cd C:\llm-test

Create the 4 task files. Run these one line at a time; each opens Notepad: paste the prompt, edit it to use your real stuff, then save and close (Ctrl+S, then Alt+F4).

notepad t1.txt
notepad t2.txt
notepad t3.txt
notepad t4.txt

(Click Yes when Notepad asks to create the file.)

Suggested prompts (edit freely):

t1 · Email rewrite

Rewrite this message so it's clear, friendly and short (Australian English). Keep all the facts. Message:
<paste a real messy email or text of yours here>

t2 · Summary

Summarise the text below in 5 bullet points, then list any action items with who and when. Text:
<paste a page of real notes, e.g. strata minutes>

t3 · PowerShell/code

Write a PowerShell 5 script that finds all files over 500MB in my Downloads folder and lists them by size, biggest first, with sizes in GB. Explain each line briefly.

t4 · Plan

Break this task into 5 small steps of 15 minutes or less, each with a clear "done when": <a real task you've been putting off>

• t1.txt to t4.txt saved in C:\llm-test

Step 4 · Run both models on all 4 tasks (⏱ 0:25–0:55)

Copy this whole block into PowerShell in one go and press Enter. It runs 8 jobs (one model at a time) and takes roughly 10–25 min. Take a break while it runs.

cd C:\llm-test
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
$models = "qwen3.8:27b", "deepseek-r1:32b"
foreach ($m in $models) {
  foreach ($t in 1..4) {
    $p = Get-Content "t$t.txt" -Raw -Encoding UTF8
    $name = ($m -replace ':', '_') + "-t$t.md"
    Write-Host "Running $m on task $t ..." -ForegroundColor Cyan
    $secs = (Measure-Command {
      ollama run $m $p | Out-File $name -Encoding utf8
    }).TotalSeconds
    $secs = [math]::Round($secs)
    Write-Host "  done in $secs s -> $name" -ForegroundColor Green
    "$m,task$t,$secs" | Out-File times.csv -Append -Encoding utf8
  }
}
Write-Host "ALL DONE" -ForegroundColor Yellow

You should see 8 cyan "Running..." lines, each followed by a green "done in ... s" line, and then ALL DONE. Check the files:

dir *.md
Get-Content times.csv

You should see 8 .md files and 8 lines of times.

• ALL DONE and 8 answer files

Step 5 · Read and score (⏱ 0:55–1:10)

Open each pair side by side, for example:

notepad qwen3.8_27b-t1.md
notepad deepseek-r1_32b-t1.md

DeepSeek shows its thinking first (Thinking... then ...done thinking.). Judge only the final answer. Score quality 1–5 (5 = I'd use it as-is). Take the times from times.csv.

• Table filled in

Step 6 · Decide (⏱ 1:10–1:15)

Write one line: "For everyday tasks I'll use because ."

(Optional: "For I'll use the other one.")

• Decision written

✅ Pass check: lesson complete when ALL are ticked

• ollama ps showed 100% GPU for both models
• Scoring table filled in for all 4 tasks
• Decision line written
• Told Skills Coach your winner and scores, so 4.1 gets marked done

Troubleshooting

Quick quiz