Skip to content

Windows validation: October 10, 2026

Local CLI and API inference passed on the machine below using Gemma 3 1B. This is a partial validation of the Windows workflow, not a completed test of the walkthrough's llama3.2:1b model or a fresh installation.

Tested environment

  • Windows 11 Home, build 10.0.26300.
  • Windows PowerShell 5.1.26100.9444.
  • Ollama client and server 0.40.2.
  • Intel Core Ultra 9 288V; approximately 31.6 GiB usable system RAM.
  • Intel Arc 140V integrated GPU present; Ollama reported 100% CPU placement for these tests, so the timings below are CPU inference. (A stray CPU-only test server was holding Ollama's port during the run; the normal setup uses the Arc GPU.)
  • gemma3:1b, Q4_K_M, 815,319,323 bytes on disk, model digest 83be9dbf9dd6ea077de84d4fbaeb094dae9f266ca1b302f04e1c803bea566a79.

What passed

  • Version query and installed-model listing.
  • ollama run gemma3:1b with the walkthrough's two-sentence prompt, supplied as a command argument; readable output appeared.
  • PowerShell Invoke-RestMethod chat requests returned nonempty greetings and done: true in all six measured requests.
  • ollama ps reported the requested context and CPU placement.
  • ollama stop gemma3:1b unloaded the model. Unloading is asynchronous: an immediate check initially failed, but a subsequent check was empty. The repeatable script waits up to 15 seconds and passed.
  • The Ollama guide's /api/embed request with the already installed nomic-embed-text returned a vector of 768 elements. This separate smoke check is not included in the chat measurement JSON.

Measurements

The prompt was “Reply with a short greeting.” Each context setting had one request after unloading, followed by a second request while loaded. Requests used num_predict: 64 and temperature: 0.2.

Requested context Ollama reported loaded size Combined process working set, second request First request Second request
4,096 880 MB 907.5 MiB 1,517 ms 125 ms
8,192 958 MB 934.8 MiB 1,546 ms 104 ms
32,768 1.0 GB 1,032.4 MiB 1,568 ms 271 ms

The working set sums ollama and llama-server processes, including the service and inference worker. It is an after-response snapshot, not peak allocation or dedicated VRAM. Ollama's loaded-size estimate and OS working set measure different things and use different units. The baseline combined working set after unloading was about 39.8 MiB.

These short prompts do not fill the context window. They test context configuration and short-answer operation, not long-document accuracy, full-context memory requirements, or sustained throughput. Responses contained 3–11 generated tokens; timing differences also reflect varying output lengths. Two requests per setting are insufficient for a general performance claim. Other apps were running.

What remains unverified

  • Fresh installation, PATH changes, interactive follow-ups and /bye.
  • Download and execution of llama3.2:1b: its download stalled and was interrupted. The existing Gemma model was used instead.
  • The walkthrough's exact API block with Llama, and the larger Llama/Qwen chat and thinking examples in the in-depth guide.
  • Open WebUI installation: Docker client 29.8.2 was installed, but its desktop engine was unavailable. No container installation was attempted.
  • Hardware recommendations for 8/16 GB machines, larger models, GPU offloading, full-context workloads, concurrency, or offline isolation.

Reproduce

From the repository directory in Windows PowerShell:

.\scripts\validate-windows.ps1 -Model gemma3:1b -OutputPath gemma3-validation.json

Omit -Model to test llama3.2:1b. The script downloads the selected model only if absent, runs local inference, and stops that model afterward. Avoid running it against a model someone else is currently using. Existing downloads remain on disk. Inspect results before sharing: the report includes machine specifications and model output.

Recorded chat measurements include the version, exact model metadata, API results, durations, placement, and memory snapshots. No general claim of AI accuracy follows from these checks.