Qwen 3.5 · 0.8B
- Cold workflow
- 51.01 sec
- Output speed
- 7.47 tokens/sec
- Validated structured run
- 28.50 sec
- Context limit
- 4,096 tokens
Case study · observed July 13, 2026
Useful local AI on a budget laptop—with honest limits.
We tested Ollama, an isolated Hermes container, and two sizes of the same open model on a 16 GB Ryzen laptop. The goal was not to win a benchmark. It was to find the smallest setup that could complete one safe, reviewable workflow.
ProcessorRyzen 5 7535HS
Usable memory13.3 GB
GraphicsRadeon 660M · 2 GB
Operating system64-bit Windows
Measured results
The smaller profile finished. The larger one did not.
Qwen 3.5 · 4B + Hermes
- Response after 200+ sec
- Incomplete
- Inference
- CPU-only
- Initial context
- 65,536 tokens
- Tuned context
- 16,384 tokens
These are observations from this one computer, not universal performance claims. Background load, model version, drivers, context, and prompt length can change the result.
What the test changed
Fit the system to the hardware.
Do not force the large agent
The full Hermes tool prompt and 4B model were not practical for this CPU-only laptop during the test.
Use a bounded workflow
The smaller model received one approved JSON file, returned a structured decision, and had no browser, email, payment, or publishing access.
Keep a human approval gate
The model produced one misleading driver statement. The validator caught missing structure; human review corrected the factual meaning.
Do not hide the driver issue
Ollama reported that the installed AMD driver was too old for GPU inference. We did not attempt a system-wide driver update during this test.
Approved output
The report after human review
The local model proposes; the operating rules and a person decide what is safe and accurate enough to use.
The 0.8B model completed the bounded workflow locally. The 4B CPU-only test did not finish within 200 seconds and was stopped.
Turn owner-approved project notes from one dedicated folder into a prioritized action plan. Never change the source notes.
Exclude email, browser profiles, passwords, payment data, unrelated folders, and automatic publishing.
Inference remained CPU-only, responses took tens of seconds, and factual review remained mandatory.
What the customer pays for
Judgment, testing, boundaries, and a usable handoff.
The open-source model costs nothing to download. The setup work is determining what fits, rejecting what does not, restricting access, validating the output, and documenting how to operate and stop the system.
Software: Ollama 0.31.2, Qwen 3.5 0.8B and 4B, Hermes Agent 0.18.2 in Docker.
References: Ollama model library · Ollama hardware support · test script, synthetic input, raw output, reviewed output, and metrics.
Cost statement: $0 in model/API charges for the test. The laptop, storage, electricity, internet connection, and setup labor remain real costs.