1. Download and extract Ollama

Our controlled setup uses Ollama 0.34.2 from the official release. Download ollama-windows-amd64.zip for a 64-bit x86 Windows machine and extract it into a local folder. This is the standalone archive route; it is different from running the graphical OllamaSetup.exe installer. ARM Windows needs a different build.

The archive we downloaded was 1,460,928,014 bytes. Its SHA-256 matched the release’s published digest, and Windows reported a valid Ollama Inc. signature on ollama.exe. Open PowerShell in the folder containing the downloaded ZIP, then check it with:

Get-FileHash -Algorithm SHA256 -LiteralPath '.\ollama-windows-amd64.zip'

For that exact 0.34.2 archive, the expected SHA-256 is:

8f3fd071a2a2f9497b562f43502c77c2b701a99d1ee5dfda28da8c786373063b

A different release has a different checksum. Compare it with its own official release record; do not use this hash to approve an unrelated download.

2. Start a local server

Open PowerShell in the extracted folder, where ollama.exe is visible. Use this first window for the server:

$env:OLLAMA_MODELS = Join-Path (Get-Location) 'models'
$env:OLLAMA_HOST = '127.0.0.1:11434'
$env:OLLAMA_NO_CLOUD = '1'
$env:OLLAMA_NUM_PARALLEL = '1'
$env:OLLAMA_MAX_LOADED_MODELS = '1'
$env:OLLAMA_VULKAN = '0'
.\ollama.exe serve

Leave that window open. The local address connects programs on this computer to Ollama. The model directory is the models folder inside your chosen runtime folder. These environment settings belong to this PowerShell process; repeat them when starting a new server session. See Ollama's configuration FAQ.

The commands limit the server to one loaded model and one parallel request. We also disabled the Vulkan backend for our measurement. That switch alone does not disable every other GPU backend; check actual placement with ollama ps below. If your goal is accelerated inference, use the GPU-specific setup documentation for your hardware rather than treating our CPU result as its expected speed.

If another Ollama server is already using port 11434, do not start a second copy or stop an unrelated process. Decide which existing setup you intend to use and check its configuration first. The standalone route does not require installing a Windows service.

3. Download and inspect the exact model

Open a second PowerShell window in the same extracted folder, then run:

$env:OLLAMA_HOST = '127.0.0.1:11434'
.\ollama.exe pull qwen3.5:9b
.\ollama.exe list
.\ollama.exe show qwen3.5:9b

Keep the size suffix: qwen3.5:9b selects the native library package used here. Do not substitute a cloud-tagged model or a separately packaged GGUF and assume the same instructions apply. Wait for the pull to finish before sending the first prompt.

Our completed download resolved to model ID 6488c96fa5fa, with qwen35 architecture and Q4_K_M quantization. ollama list and ollama show both completed successfully. The full manifest digest is in the downloadable record:

6488c96fa5faab64bb65cbd30d4289e20e6130ef535a93ef9a49f42eda893ea7

In our test, the package was downloaded through Ollama’s native pull API. The CLI pull syntax above was checked against the installed runtime’s help; the inspection commands were executed. A model tag can later resolve to a different package, so compare the recorded identity when repeating the test.

Keep the server running. In the second PowerShell window, send a short, bounded request:

$request = @{
    model = 'qwen3.5:9b'
    prompt = 'Reply with exactly: Local model ready.'
    think = $false
    stream = $false
    keep_alive = '5m'
    options = @{
        num_ctx = 4096
        num_predict = 128
        num_thread = 8
        temperature = 0
        seed = 42
    }
} | ConvertTo-Json -Depth 5

$answer = Invoke-RestMethod -Uri 'http://127.0.0.1:11434/api/generate' `
    -Method Post -ContentType 'application/json' -Body $request -TimeoutSec 180
$answer.response

This uses Ollama’s local generation API. think=false disables thinking mode for this demonstration. num_predict=128 caps the generated output, while num_ctx=4096 sets a separate context window. They are different limits. Eight inference threads are our chosen setting for the active desktop, not an optimal value for every CPU.

The PowerShell request returned this exact answer:

Local model ready.

Run the block directly in PowerShell. This nonstreaming example returns the completed answer; it does not measure time to first token.

To ask your own question, replace the prompt value and resend the request. Start with a short task, such as extracting fields from a paragraph you wrote. Each example here sends a fresh prompt without conversation history; keeping the model loaded does not add previous questions to that request.

Then inspect where the model is running:

.\ollama.exe ps

Read the actual PROCESSOR and context columns. A detected graphics adapter is not proof that inference is using it. GPU behavior depends on the backend and supported hardware; consult the current GPU documentation for your device.

When you finish, use the second PowerShell window in the extracted runtime folder to unload this model and check the loaded-model list:

.\ollama.exe stop qwen3.5:9b
.\ollama.exe ps

Check that qwen3.5:9b no longer appears in the ps output. Unloading keeps the downloaded files so you can use the model again. In the first PowerShell window, press Ctrl+C to stop the foreground server you started. A separately installed background Ollama application is a different process; closing a command window does not necessarily stop that application.

In our test, this sequence unloaded the model: ollama ps displayed only its header, and the local loaded-model API returned an empty list.

The model files for the setup above are in your chosen runtime folder’s models directory. Do not delete other models or a general Windows temporary directory to free space. If you later want to remove this specific downloaded model, use Ollama’s model-management commands after checking the selected model name.

If the pull repeatedly reports stalled parts, Ollama’s downloader supports resuming from saved partial progress. Leave the server window open, cancel only the pull command with Ctrl+C, allow it to stop, then rerun the same pull command against the same model directory. Keep the partial files. Resume support does not guarantee that every in-flight byte will be retained. See the 0.34.2 downloader implementation.

If startup fails with a profile-directory error, read the message in the server window and check the location against Ollama’s Windows documentation. Our test used a separate project profile; that was a condition of the test environment, not a required step for this walkthrough.

Does local Qwen require a paid API?

This local downloaded-model route does not use a hosted inference API or need a hosted API key. Downloads, disk space, electricity and computer time still have costs. For provider-hosted access instead, use the separate Qwen API guide.

Is this the same as proving it works offline?

No. The setup uses a local model and disables Ollama cloud features, but we do not disconnect the computer from the network. Downloads need internet access. A network-isolation test and a review of any extra frontend or tool are separate checks; do not infer a universal privacy guarantee from the word “local.”

Can I connect this to Qwen Code?

A working local runtime is only one prerequisite. Our Qwen Code guide explains the coding-agent client and its separate provider-backed test. This 4K-context text experiment does not validate a local Qwen Code connection, tool use or a long-context repository task.

The local workflow ran successfully, but only the extraction task met every answer criterion. The summary was incomplete and the arithmetic answer was wrong in each repetition. Successful completion and a short wait are separate from answer quality.

These are three distinct tasks, each repeated three times with the same settings. They do not establish general accuracy or predict performance on a 16 GB laptop, large documents, vision tasks or coding-agent work.

These are observations from a machine with unusually large RAM capacity, not minimum hardware requirements. Our sampling settings were fixed for this small experiment; we did not optimize them or compare thinking modes.

For example, the resident extraction requests started showing text after about 0.26–0.27 seconds, but took 16.36–16.54 seconds to finish the JSON. Generation rate for the longer extraction and summary responses ranged from 4.92 to 4.98 tokens/second. The raw CSV includes load time, token counts, generation rates and memory figures for every request.

The 6.55 GiB figure above is a sampled sum of process working sets. It can count shared pages twice and miss a peak, so it is not a minimum-RAM requirement. The desktop was in active use, and repeated prompts may benefit from caching as well as an already loaded model.

Read the exact prompts and unedited answers

Exact prompts and first answers

The first answer for each task is reproduced below. Both resident repetitions produced the same text as that task’s first response.

JSON extraction

Exact prompt:

Extract only the facts below. Return one JSON object with exactly these keys: event, date, time, city, capacity, ticket_gbp, contact_email. Use strings for event/date/time/city, numbers for capacity/ticket_gbp, and null for missing information. No Markdown or additional text.

Fictional event: Harbor Code Clinic.
Date: 14 October 2026. Start time: 18:30.
City: Bristol. Maximum attendance: 24 people.
Standard ticket: GBP 12. No contact email has been published.
Use YYYY-MM-DD for date and HH:MM for time.

First unedited answer:

{
  "event": "Harbor Code Clinic",
  "date": "2026-10-14",
  "time": "18:30",
  "city": "Bristol",
  "capacity": 24,
  "ticket_gbp": 12,
  "contact_email": null
}

Result: met the criterion in all three runs. The JSON had exactly the required fields, values and types, including null for the missing email. We checked valid JSON and duplicate keys as well as the extracted facts.

Summary

Exact prompt:

Summarize this fictional update in exactly three bullets, beginning each with "- ". Use one bullet each for opening, completed work, and remaining restrictions. Each bullet must contain at most 25 words. Include the stated numbers and preserve uncertainty. Add no facts.

Riverside Pocket Park opens on 4 October 2026, with daily opening hours of 07:00–21:00. Phase one added a 2 km walking path and 30 trees. The playground remains closed pending inspection; no reopening date has been announced. Drinking fountains are not installed, so visitors should bring water.

First unedited answer:

- Riverside Pocket Park opens October 4, 2026, with daily hours from 07:00 to 21:00.
- Phase one added a 2 km walking path and 30 trees.
- The playground remains closed pending inspection; no reopening date has been announced. Drinking fountains are not installed.

Result: incomplete in all three runs. It omitted the source’s advice to bring water, so it failed our predeclared completeness criterion despite preserving the other facts and required format.

Arithmetic

Exact prompt:

A fictional shop starts with 12 boxes containing 18 notebooks each. It sells 47 notebooks, donates 13, and then receives 20 additional notebooks. How many notebooks remain?

Return only the count as a base-10 integer, without units or explanation.

First unedited answer:

59

Result: wrong in all three runs. The expected count is 176: 12 × 18 − 47 − 13 + 20. The model returned 59. This is one short, fixed non-thinking task; it does not establish a general mathematics score. Verify numbers before using them.

Read the complete records as plain text (TXT, 187 KB), or download the original experiment records and method notes (ZIP, 48 KB). The archive includes the frozen prompts and criteria, all nine raw responses, two readiness checks, model identity, per-request metrics, answer review and walkthrough verification. Start with README.md; the included measurement code is for inspection and requires the original project layout to run. The archive contains no model binaries or API credentials. A private script-file path in the blocked-launch record is explicitly redacted; the model answers are unedited.

For a ready-to-run check of the saved results, download the standalone offline evidence checker (ZIP, 32 KB). Extract it and run python check_saved_qwen.py with Python 3.9 or later. It checks the nine saved requests, original mechanical scores and timing calculations using only Python’s standard library. It makes no model calls. A successful record check means the files are consistent, not that every answer was correct; summary meaning still needs the included written review.